Dataset Viewer
topic
stringclasses 1
value | RAG
stringclasses 1
value | prompt
stringlengths 15.5k
19.3k
|
|---|---|---|
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
abstract: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
Speech language models (SpeechLMs) accept speech input and produce speech
output, allowing for more natural human-computer interaction compared to text-
based large language models (LLMs). Traditional approaches for developing
SpeechLMs are constrained by the limited availability of unsupervised speech
data and parallel speech-text data, which are significantly less abundant than text
pre-training data, thereby limiting their scalability as LLMs. We propose a novel
approach to scaling speech-text pre-training by leveraging large-scale synthetic in-
terleaved data derived from text corpora, eliminating the need for parallel speech-
text datasets. Our method efficiently constructs speech-text interleaved data by
sampling text spans from existing text corpora and synthesizing corresponding
speech spans using a text-to-token model, bypassing the need to generate actual
speech. We also employ a supervised speech tokenizer derived from an auto-
matic speech recognition (ASR) model by incorporating a vector-quantized bot-
tleneck into the encoder. This supervised training approach results in discrete
speech tokens with strong semantic preservation even at lower frame rates (e.g.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
abstract: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
Training on high-quality synthetic data from strong language models (LMs) is a
common strategy to improve the reasoning performance of LMs. In this work,
we revisit whether this strategy is compute-optimal under a fixed inference bud-
get (e.g., FLOPs). To do so, we investigate the trade-offs between generating
synthetic data using a stronger but more expensive (SE) model versus a weaker
but cheaper (WC) model. We evaluate the generated data across three key met-
rics: coverage, diversity, and false positive rate, and show that the data from WC
models may have higher coverage and diversity, but also exhibit higher false pos-
itive rates. We then finetune LMs on data from SE and WC models in different
settings: knowledge distillation, self-improvement, and a novel weak-to-strong
improvement setup where a weaker LM teaches reasoning to a stronger LM. Our
findings reveal that models finetuned on WC-generated data consistently outper-
form those trained on SE-generated data across multiple benchmarks and multiple
choices of WC and SE models. These results challenge the prevailing practice of
relying on SE models for synthetic data generation, suggesting that WC may be
the compute-optimal approach for training advanced LM reasoners.
(a) Finetuning LMs with Gemma2 data.
(b) Finetuning LMs with Gemini 1.5 data.
Figure 1: Summary of the results. (a) We finetune Gemma-7B, Gemma2-9B, and Gemma2-27B on
the synthetic data collected from a stronger but more expensive LM (Gemma2-27B) and a weaker
but cheaper LM (Gemma2-9B) in a compute-matched setup for the MATH dataset. We find that
training with Gemma2-9B data is more compute-optimal across diverse finetuning paradigms –
knowledge distillation, self-improvement, and weak-to-strong improvement (i.e. using a weaker
model to improve a stronger model). (b) We finetune Gemma models (7B/9B/27B) on synthetic
data generated by Gemini-1.5-Pro and Gemini-1.5-Flash in a price-matched setup. We find that
finetuning with Flash-generated data consistently outperforms Pro-generated data.
∗Llama experiments in this paper were conducted only by parties outside of Google. Authors affiliated with
Google were responsible for the Gemini and Gemma experiments.
title: Synthetic continued pretraining
abstract: Synthetic continued pretraining
Pretraining on large-scale, unstructured internet text enables language models to
acquire a significant amount of world knowledge. However, this knowledge acqui-
sition is data-inefficient—to learn a fact, models must be trained on hundreds to
thousands of diverse representations of it. This poses a challenge when adapting a
pretrained model to a small corpus of domain-specific documents, where each fact
may appear rarely or only once. We propose to bridge this gap with synthetic con-
tinued pretraining: using the small domain-specific corpus to synthesize a large
corpus more amenable to learning, and then performing continued pretraining on
the synthesized corpus. We instantiate this proposal with EntiGraph, a synthetic
data augmentation algorithm that extracts salient entities from the source corpus
and then generates diverse text by drawing connections between those entities.
Synthetic continued pretraining with EntiGraph enables a language model to an-
swer questions and follow generic instructions related to the source documents
without access to them. If the source documents are instead available at inference
time, we show that the knowledge acquired through our approach compounds with
retrieval-augmented generation. To better understand these results, we build a sim-
ple mathematical model of EntiGraph, and show how synthetic data augmentation
can “rearrange” knowledge to enable more data-efficient learning.
title: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
abstract: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
The utility of synthetic data to enhance pretraining data quality and hence to im-
prove downstream task accuracy has been widely explored in recent large lan-
guage models (LLMs). Yet, these approaches fall inadequate in complex, multi-
hop and mathematical reasoning tasks as the synthetic data typically fails to add
complementary knowledge to the existing raw corpus. In this work, we propose a
novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) gener-
ation method that improves the mathematical reasoning ability of LLMs. Specifi-
cally, using MIND, we generate synthetic conversations based on OpenWebMath
(OWM), resulting in a new math corpus, MIND-OWM. Our experiments with dif-
ferent conversational settings reveal that incorporating knowledge gaps between
dialog participants is essential for generating high-quality math data. We further
identify an effective way to format and integrate synthetic and raw data during pre-
training to maximize the gain in mathematical reasoning, emphasizing the need to
restructure raw data rather than use it as-is. Compared to pretraining just on raw
data, a model pretrained on MIND-OWM shows significant boost in mathematical
reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance
in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general
purpose reasoning tasks (GENERAL REASONING: +2.51%).
title: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
abstract: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
Software engineers mainly write code by editing existing programs. In contrast,
language models (LMs) autoregressively synthesize programs in a single pass.
One explanation for this is the scarcity of sequential edit data. While high-quality
instruction data for code synthesis is scarce, edit data for synthesis is even scarcer.
To fill this gap, we develop a synthetic data generation algorithm called LintSeq.
This algorithm refactors programs into sequences of synthetic edits by using a
linter to procedurally sample across interdependent lines of source code. Synthetic
edits sampled with LintSeq reflect the syntax and semantics of their programming
language. To test the algorithm, we use it to refactor a dataset of instruction +
program pairs into instruction + program-diff-sequence tuples. Then, we fine-
tune a series of smaller LMs ranging from 2.6B to 14B parameters on both the
re-factored and original versions of this dataset. We perform comprehensive
evaluations comparing edit sequence code LMs against baselines on HumanEval,
MBPP(+), CodeContests, DS-1000, and BigCodeBench. We show that models
fine-tuned to iteratively synthesize code match or outperform baselines on pass@1,
and exhibit better scaling across higher pass@k as a function of total test-time
FLOPs. Finally, we also pretrain our own tiny LMs for code understanding. We
show that fine-tuning these models to synthesize code edit-by-edit results in strong
performance on HumanEval and MBPP(+) compared to existing code language
models of similar scale such as CodeT5+, AlphaCode, and Codex.
title: DataGen: Unified Synthetic Dataset Generation via Large Language Models
abstract: DataGen: Unified Synthetic Dataset Generation via Large Language Models
Large Language Models (LLMs) such as GPT-4 and Llama3 have significantly
impacted various fields by enabling high-quality synthetic data generation and
reducing dependence on expensive human-generated datasets. Despite this, chal-
lenges remain in the areas of generalization, controllability, diversity, and truthful-
ness within the existing generative frameworks. To address these challenges, this
paper presents DATAGEN, a comprehensive LLM-powered framework designed to
produce diverse, accurate, and highly controllable datasets. DATAGEN is adaptable,
supporting all types of text datasets and enhancing the generative process through
innovative mechanisms. To augment data diversity, DATAGEN incorporates an
attribute-guided generation module and a group checking feature. For accuracy, it
employs a code-based mathematical assessment for label verification alongside a
retrieval-augmented generation technique for factual validation. The framework
also allows for user-specified constraints, enabling customization of the data gener-
ation process to suit particular requirements. Extensive experiments demonstrate
the superior quality of data generated by DATAGEN, and each module within
DATAGEN plays a critical role in this enhancement. Additionally, DATAGEN is
applied in two practical scenarios: benchmarking LLMs and data augmentation.
The results indicate that DATAGEN effectively supports dynamic and evolving
benchmarking and that data augmentation improves LLM capabilities in various
domains, including agent-oriented abilities and reasoning skills.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
abstract: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
Biological language model performance depends heavily on pretraining data
quality, diversity, and size. While metagenomic datasets feature enor-
mous biological diversity, their utilization as pretraining data has been
limited due to challenges in data accessibility, quality filtering and dedupli-
cation. Here, we present the Open MetaGenomic (OMG) corpus, a genomic
pretraining dataset totalling 3.1T base pairs and 3.3B protein coding se-
quences, obtained by combining two largest metagenomic dataset reposito-
ries (JGI’s IMG and EMBL’s MGnify). We first document the composition
of the dataset and describe the quality filtering steps taken to remove poor
quality data. We make the OMG corpus available as a mixed-modality
genomic sequence dataset that represents multi-gene encoding genomic
sequences with translated amino acids for protein coding sequences, and
nucleic acids for intergenic sequences. We train the first mixed-modality
genomic language model (gLM2) that leverages genomic context informa-
tion to learn robust functional representations, as well as coevolutionary
signals in protein-protein interfaces and genomic regulatory syntax. Fur-
thermore, we show that deduplication in embedding space can be used to
balance the corpus, demonstrating improved performance on downstream
tasks. The OMG dataset is publicly hosted on the Hugging Face Hub
at https://huggingface.co/datasets/tattabio/OMG and gLM2 is avail-
able at https://huggingface.co/tattabio/gLM2_650M.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: Self-Boosting Large Language Models with Synthetic Preference Data
abstract: Self-Boosting Large Language Models with Synthetic Preference Data
Through alignment with human preferences, Large Language Models (LLMs)
have advanced significantly in generating honest, harmless, and helpful responses.
However, collecting high-quality preference data is a resource-intensive and
creativity-demanding process, especially for the continual improvement of LLMs.
We introduce SynPO, a self-boosting paradigm that leverages synthetic prefer-
ence data for model alignment. SynPO employs an iterative mechanism wherein
a self-prompt generator creates diverse prompts, and a response improver refines
model responses progressively. This approach trains LLMs to autonomously learn
the generative rewards for their own outputs and eliminates the need for large-
scale annotation of prompts and human preferences. After four SynPO itera-
tions, Llama3-8B and Mistral-7B show significant enhancements in instruction-
following abilities, achieving over 22.1% win rate improvements on AlpacaEval
2.0 and ArenaHard. Simultaneously, SynPO improves the general performance
of LLMs on various tasks, validated by a 3.2 to 5.0 average score increase on the
well-recognized Open LLM leaderboard.
title: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
abstract: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
The utility of synthetic data to enhance pretraining data quality and hence to im-
prove downstream task accuracy has been widely explored in recent large lan-
guage models (LLMs). Yet, these approaches fall inadequate in complex, multi-
hop and mathematical reasoning tasks as the synthetic data typically fails to add
complementary knowledge to the existing raw corpus. In this work, we propose a
novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) gener-
ation method that improves the mathematical reasoning ability of LLMs. Specifi-
cally, using MIND, we generate synthetic conversations based on OpenWebMath
(OWM), resulting in a new math corpus, MIND-OWM. Our experiments with dif-
ferent conversational settings reveal that incorporating knowledge gaps between
dialog participants is essential for generating high-quality math data. We further
identify an effective way to format and integrate synthetic and raw data during pre-
training to maximize the gain in mathematical reasoning, emphasizing the need to
restructure raw data rather than use it as-is. Compared to pretraining just on raw
data, a model pretrained on MIND-OWM shows significant boost in mathematical
reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance
in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general
purpose reasoning tasks (GENERAL REASONING: +2.51%).
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
abstract: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
Training on high-quality synthetic data from strong language models (LMs) is a
common strategy to improve the reasoning performance of LMs. In this work,
we revisit whether this strategy is compute-optimal under a fixed inference bud-
get (e.g., FLOPs). To do so, we investigate the trade-offs between generating
synthetic data using a stronger but more expensive (SE) model versus a weaker
but cheaper (WC) model. We evaluate the generated data across three key met-
rics: coverage, diversity, and false positive rate, and show that the data from WC
models may have higher coverage and diversity, but also exhibit higher false pos-
itive rates. We then finetune LMs on data from SE and WC models in different
settings: knowledge distillation, self-improvement, and a novel weak-to-strong
improvement setup where a weaker LM teaches reasoning to a stronger LM. Our
findings reveal that models finetuned on WC-generated data consistently outper-
form those trained on SE-generated data across multiple benchmarks and multiple
choices of WC and SE models. These results challenge the prevailing practice of
relying on SE models for synthetic data generation, suggesting that WC may be
the compute-optimal approach for training advanced LM reasoners.
(a) Finetuning LMs with Gemma2 data.
(b) Finetuning LMs with Gemini 1.5 data.
Figure 1: Summary of the results. (a) We finetune Gemma-7B, Gemma2-9B, and Gemma2-27B on
the synthetic data collected from a stronger but more expensive LM (Gemma2-27B) and a weaker
but cheaper LM (Gemma2-9B) in a compute-matched setup for the MATH dataset. We find that
training with Gemma2-9B data is more compute-optimal across diverse finetuning paradigms –
knowledge distillation, self-improvement, and weak-to-strong improvement (i.e. using a weaker
model to improve a stronger model). (b) We finetune Gemma models (7B/9B/27B) on synthetic
data generated by Gemini-1.5-Pro and Gemini-1.5-Flash in a price-matched setup. We find that
finetuning with Flash-generated data consistently outperforms Pro-generated data.
∗Llama experiments in this paper were conducted only by parties outside of Google. Authors affiliated with
Google were responsible for the Gemini and Gemma experiments.
title: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
abstract: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
Synthetic data augmentation via Large Language Models (LLMs) allows re-
searchers to leverage additional training data, thus enhancing the performance of
downstream tasks, especially when real-world data is scarce. However, the gen-
erated data can deviate from the real-world data, and this misalignment can bring
about deficient results while applying the trained model to applications. There-
fore, we proposed efficient weighted-loss approaches to align synthetic data with
real-world distribution by emphasizing high-quality and diversified data generated
by LLMs using merely a tiny amount of real-world data. We empirically assessed
the effectiveness of our methods on multiple text classification tasks, and the re-
sults showed that leveraging our approaches on a BERT-level model robustly out-
performed standard cross-entropy and other data weighting approaches, providing
potential solutions to effectively leveraging synthetic data from any suitable data
generator.
title: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
abstract: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
Biological language model performance depends heavily on pretraining data
quality, diversity, and size. While metagenomic datasets feature enor-
mous biological diversity, their utilization as pretraining data has been
limited due to challenges in data accessibility, quality filtering and dedupli-
cation. Here, we present the Open MetaGenomic (OMG) corpus, a genomic
pretraining dataset totalling 3.1T base pairs and 3.3B protein coding se-
quences, obtained by combining two largest metagenomic dataset reposito-
ries (JGI’s IMG and EMBL’s MGnify). We first document the composition
of the dataset and describe the quality filtering steps taken to remove poor
quality data. We make the OMG corpus available as a mixed-modality
genomic sequence dataset that represents multi-gene encoding genomic
sequences with translated amino acids for protein coding sequences, and
nucleic acids for intergenic sequences. We train the first mixed-modality
genomic language model (gLM2) that leverages genomic context informa-
tion to learn robust functional representations, as well as coevolutionary
signals in protein-protein interfaces and genomic regulatory syntax. Fur-
thermore, we show that deduplication in embedding space can be used to
balance the corpus, demonstrating improved performance on downstream
tasks. The OMG dataset is publicly hosted on the Hugging Face Hub
at https://huggingface.co/datasets/tattabio/OMG and gLM2 is avail-
able at https://huggingface.co/tattabio/gLM2_650M.
title: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
abstract: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
Synthetic data has been widely used to train large language models, but their gener-
ative nature inevitably introduces noisy, non-informative, and misleading learning
signals. In this paper, we propose MONTESSORI-INSTRUCT, a novel data synthesis
framework that tailors the data synthesis ability of the teacher language model
toward the student language model’s learning process. Specifically, we utilize local
data influence of synthetic training data points on students to characterize students’
learning preferences. Then, we train the teacher model with Direct Preference
Optimization (DPO) to generate synthetic data tailored toward student learning pref-
erences. Experiments with Llama3-8B-Instruct (teacher) and Llama3-8B (student)
on Alpaca Eval and MT-Bench demonstrate that Montessori-Instruct significantly
outperforms standard synthesis methods by 18.35% and 46.24% relatively. Our
method also beats data synthesized by a stronger teacher model, GPT-4o. Further
analysis confirms the benefits of teacher’s learning to generate more influential train-
ing data in the student’s improved learning, the advantages of local data influence
in accurately measuring student preferences, and the robustness of Montessori-
Instruct across different student models. Our code and data are open-sourced at
https://github.com/cxcscmu/Montessori-Instruct.
title: Strong Model Collapse
abstract: Strong Model Collapse
Within the scaling laws paradigm, which underpins the training of large neural
networks like ChatGPT and Llama, we consider a supervised regression setting
and establish a strong form of the model collapse phenomenon, a critical perfor-
mance degradation due to synthetic data in the training corpus. Our results show
that even the smallest fraction of synthetic data (e.g., as little as 1 per 1000) can
still lead to model collapse: larger and larger training sets do not enhance perfor-
mance. We further investigate whether increasing model size, an approach aligned
with current trends in training large language models, exacerbates or mitigates
model collapse. In a simplified regime where neural networks are approximated
via random projections of tunable size, we both theoretically and empirically show
that larger models can amplify model collapse. Interestingly, our theory also in-
dicates that, beyond the interpolation threshold (which can be extremely high for
very large datasets), larger models may mitigate the collapse, although they do
not entirely prevent it. Our theoretical findings are empirically verified through
experiments on language models and neural networks for images.
title: Synthetic continued pretraining
abstract: Synthetic continued pretraining
Pretraining on large-scale, unstructured internet text enables language models to
acquire a significant amount of world knowledge. However, this knowledge acqui-
sition is data-inefficient—to learn a fact, models must be trained on hundreds to
thousands of diverse representations of it. This poses a challenge when adapting a
pretrained model to a small corpus of domain-specific documents, where each fact
may appear rarely or only once. We propose to bridge this gap with synthetic con-
tinued pretraining: using the small domain-specific corpus to synthesize a large
corpus more amenable to learning, and then performing continued pretraining on
the synthesized corpus. We instantiate this proposal with EntiGraph, a synthetic
data augmentation algorithm that extracts salient entities from the source corpus
and then generates diverse text by drawing connections between those entities.
Synthetic continued pretraining with EntiGraph enables a language model to an-
swer questions and follow generic instructions related to the source documents
without access to them. If the source documents are instead available at inference
time, we show that the knowledge acquired through our approach compounds with
retrieval-augmented generation. To better understand these results, we build a sim-
ple mathematical model of EntiGraph, and show how synthetic data augmentation
can “rearrange” knowledge to enable more data-efficient learning.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
abstract: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
The rise of Large Language Models (LLMs) has accentuated the need for diverse,
high-quality pre-training data. Synthetic data emerges as a viable solution to the
challenges of data scarcity and inaccessibility. While previous literature has fo-
cused predominantly on the quality and quantity of real data, our work enables the
measurement of diversity in synthetic data and explores its impact on LLM perfor-
mance. We study the downstream effects of synthetic data diversity during both
the pre-training and fine-tuning stages by introducing a new diversity metric, LLM
cluster-agent, designed to evaluate the diversity of synthetic datasets. Through a
series of controlled experiments with models of 350M and 1.4B parameters, we
demonstrate that the proposed cluster-based LLM scoring of diversity correlates
positively with both pre-training and supervised fine-tuning performance. Our
findings also reveal that synthetic data diversity in pre-training affects supervised
fine-tuning more significantly than pre-training itself, even for smaller models. We
hope this study advances our understanding of the optimal use of synthetic data in
LLM training and opens new avenues for efficient data generation processes.
title: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
abstract: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
The utility of synthetic data to enhance pretraining data quality and hence to im-
prove downstream task accuracy has been widely explored in recent large lan-
guage models (LLMs). Yet, these approaches fall inadequate in complex, multi-
hop and mathematical reasoning tasks as the synthetic data typically fails to add
complementary knowledge to the existing raw corpus. In this work, we propose a
novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) gener-
ation method that improves the mathematical reasoning ability of LLMs. Specifi-
cally, using MIND, we generate synthetic conversations based on OpenWebMath
(OWM), resulting in a new math corpus, MIND-OWM. Our experiments with dif-
ferent conversational settings reveal that incorporating knowledge gaps between
dialog participants is essential for generating high-quality math data. We further
identify an effective way to format and integrate synthetic and raw data during pre-
training to maximize the gain in mathematical reasoning, emphasizing the need to
restructure raw data rather than use it as-is. Compared to pretraining just on raw
data, a model pretrained on MIND-OWM shows significant boost in mathematical
reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance
in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general
purpose reasoning tasks (GENERAL REASONING: +2.51%).
title: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
abstract: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
Recent studies have shown that Large Language Models (LLMs) struggle to accu-
rately retrieve information and maintain reasoning capabilities when processing
long-context inputs. To address these limitations, we propose a finetuning approach
utilizing a carefully designed synthetic dataset comprising numerical key-value
retrieval tasks. Our experiments on models like GPT-3.5 Turbo and Mistral 7B
demonstrate that finetuning LLMs on this dataset significantly improves LLMs’ in-
formation retrieval and reasoning capabilities in longer-context settings. We present
an analysis of the finetuned models, illustrating the transfer of skills from synthetic
to real task evaluations (e.g., 10.5% improvement on 20 documents MDQA at
position 10 for GPT-3.5 Turbo). We also find that finetuned LLMs’ performance
on general benchmarks remains almost constant while LLMs finetuned on other
baseline long-context augmentation data can encourage hallucination (e.g., on
TriviaQA, Mistral 7B finetuned on our synthetic data cause no performance drop
while other baseline data can cause a drop that ranges from 2.33% to 6.19%). Our
study highlights the potential of finetuning on synthetic data for improving the
performance of LLMs on longer-context tasks.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
abstract: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
Biological language model performance depends heavily on pretraining data
quality, diversity, and size. While metagenomic datasets feature enor-
mous biological diversity, their utilization as pretraining data has been
limited due to challenges in data accessibility, quality filtering and dedupli-
cation. Here, we present the Open MetaGenomic (OMG) corpus, a genomic
pretraining dataset totalling 3.1T base pairs and 3.3B protein coding se-
quences, obtained by combining two largest metagenomic dataset reposito-
ries (JGI’s IMG and EMBL’s MGnify). We first document the composition
of the dataset and describe the quality filtering steps taken to remove poor
quality data. We make the OMG corpus available as a mixed-modality
genomic sequence dataset that represents multi-gene encoding genomic
sequences with translated amino acids for protein coding sequences, and
nucleic acids for intergenic sequences. We train the first mixed-modality
genomic language model (gLM2) that leverages genomic context informa-
tion to learn robust functional representations, as well as coevolutionary
signals in protein-protein interfaces and genomic regulatory syntax. Fur-
thermore, we show that deduplication in embedding space can be used to
balance the corpus, demonstrating improved performance on downstream
tasks. The OMG dataset is publicly hosted on the Hugging Face Hub
at https://huggingface.co/datasets/tattabio/OMG and gLM2 is avail-
able at https://huggingface.co/tattabio/gLM2_650M.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
abstract: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
Software engineers mainly write code by editing existing programs. In contrast,
language models (LMs) autoregressively synthesize programs in a single pass.
One explanation for this is the scarcity of sequential edit data. While high-quality
instruction data for code synthesis is scarce, edit data for synthesis is even scarcer.
To fill this gap, we develop a synthetic data generation algorithm called LintSeq.
This algorithm refactors programs into sequences of synthetic edits by using a
linter to procedurally sample across interdependent lines of source code. Synthetic
edits sampled with LintSeq reflect the syntax and semantics of their programming
language. To test the algorithm, we use it to refactor a dataset of instruction +
program pairs into instruction + program-diff-sequence tuples. Then, we fine-
tune a series of smaller LMs ranging from 2.6B to 14B parameters on both the
re-factored and original versions of this dataset. We perform comprehensive
evaluations comparing edit sequence code LMs against baselines on HumanEval,
MBPP(+), CodeContests, DS-1000, and BigCodeBench. We show that models
fine-tuned to iteratively synthesize code match or outperform baselines on pass@1,
and exhibit better scaling across higher pass@k as a function of total test-time
FLOPs. Finally, we also pretrain our own tiny LMs for code understanding. We
show that fine-tuning these models to synthesize code edit-by-edit results in strong
performance on HumanEval and MBPP(+) compared to existing code language
models of similar scale such as CodeT5+, AlphaCode, and Codex.
title: DataGen: Unified Synthetic Dataset Generation via Large Language Models
abstract: DataGen: Unified Synthetic Dataset Generation via Large Language Models
Large Language Models (LLMs) such as GPT-4 and Llama3 have significantly
impacted various fields by enabling high-quality synthetic data generation and
reducing dependence on expensive human-generated datasets. Despite this, chal-
lenges remain in the areas of generalization, controllability, diversity, and truthful-
ness within the existing generative frameworks. To address these challenges, this
paper presents DATAGEN, a comprehensive LLM-powered framework designed to
produce diverse, accurate, and highly controllable datasets. DATAGEN is adaptable,
supporting all types of text datasets and enhancing the generative process through
innovative mechanisms. To augment data diversity, DATAGEN incorporates an
attribute-guided generation module and a group checking feature. For accuracy, it
employs a code-based mathematical assessment for label verification alongside a
retrieval-augmented generation technique for factual validation. The framework
also allows for user-specified constraints, enabling customization of the data gener-
ation process to suit particular requirements. Extensive experiments demonstrate
the superior quality of data generated by DATAGEN, and each module within
DATAGEN plays a critical role in this enhancement. Additionally, DATAGEN is
applied in two practical scenarios: benchmarking LLMs and data augmentation.
The results indicate that DATAGEN effectively supports dynamic and evolving
benchmarking and that data augmentation improves LLM capabilities in various
domains, including agent-oriented abilities and reasoning skills.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
abstract: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
The rise of Large Language Models (LLMs) has accentuated the need for diverse,
high-quality pre-training data. Synthetic data emerges as a viable solution to the
challenges of data scarcity and inaccessibility. While previous literature has fo-
cused predominantly on the quality and quantity of real data, our work enables the
measurement of diversity in synthetic data and explores its impact on LLM perfor-
mance. We study the downstream effects of synthetic data diversity during both
the pre-training and fine-tuning stages by introducing a new diversity metric, LLM
cluster-agent, designed to evaluate the diversity of synthetic datasets. Through a
series of controlled experiments with models of 350M and 1.4B parameters, we
demonstrate that the proposed cluster-based LLM scoring of diversity correlates
positively with both pre-training and supervised fine-tuning performance. Our
findings also reveal that synthetic data diversity in pre-training affects supervised
fine-tuning more significantly than pre-training itself, even for smaller models. We
hope this study advances our understanding of the optimal use of synthetic data in
LLM training and opens new avenues for efficient data generation processes.
title: Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective
abstract: Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective
Synthetic data has become a pivotal resource in post-training tasks for large lan-
guage models (LLMs) due to the scarcity of high-quality, specific data. While
various methods have been developed to generate synthetic data, there remains a
discernible gap between the practical effects of synthetic data and our theoretical
comprehension. To address this challenge, we commence by presenting a detailed
modeling of the prevalent synthetic data generation process. Building upon this
modeling, we demonstrate that the generalization capability of the post-trained
model is critically determined by the information gain derived from the generative
model, as analyzed from a novel reverse-bottleneck perspective. Moreover, we in-
troduce the concept of Generalization Gain via Mutual Information (GGMI) and
elucidate the relationship between generalization gain and information gain. This
analysis serves as a theoretical foundation for synthetic data generation and further
highlights its connection with the generalization capability of post-trained models,
offering an understanding about the design of synthetic data generation techniques
and the optimization of the post-training process. We open-source our code at
https://github.com/ZyGan1999/Towards-a-Theoretical-U
nderstanding-of-Synthetic-Data-in-LLM-Post-Training.
title: Self-Boosting Large Language Models with Synthetic Preference Data
abstract: Self-Boosting Large Language Models with Synthetic Preference Data
Through alignment with human preferences, Large Language Models (LLMs)
have advanced significantly in generating honest, harmless, and helpful responses.
However, collecting high-quality preference data is a resource-intensive and
creativity-demanding process, especially for the continual improvement of LLMs.
We introduce SynPO, a self-boosting paradigm that leverages synthetic prefer-
ence data for model alignment. SynPO employs an iterative mechanism wherein
a self-prompt generator creates diverse prompts, and a response improver refines
model responses progressively. This approach trains LLMs to autonomously learn
the generative rewards for their own outputs and eliminates the need for large-
scale annotation of prompts and human preferences. After four SynPO itera-
tions, Llama3-8B and Mistral-7B show significant enhancements in instruction-
following abilities, achieving over 22.1% win rate improvements on AlpacaEval
2.0 and ArenaHard. Simultaneously, SynPO improves the general performance
of LLMs on various tasks, validated by a 3.2 to 5.0 average score increase on the
well-recognized Open LLM leaderboard.
title: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
abstract: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
We explore model collapse caused by synthetic data, where AI models trained on
such data experience a gradual decline in performance. Our initial analysis exam-
ines language model pretraining on mixed human and synthetic data, highlighting
performance degradation. Further statistical analysis reveals distributional shifts
and an over-concentration of n-gram features caused by synthetic data. Inspired
by these insights, we propose token-level editing on human data, to obtain semi-
synthetic data instead of fully using model outputs. As a proof of concept, we
theoretically demonstrate that token-level editing can prevent model collapse, as
the test error is constrained by a finite upper bound. We conducted extensive ex-
periments on pretraining, continual pretraining, and supervised fine-tuning of lan-
guage models. The results validate our theoretical proof that token-level editing
improves data quality and enhances model performance.
title: Scalable Extraction of Training Data from Aligned, Production Language Models
abstract: Scalable Extraction of Training Data from Aligned, Production Language Models
Large language models are prone to memorizing some of their training data. Mem-
orized (and possibly sensitive) samples can then be extracted at generation time
by adversarial or benign users. There is hope that model alignment—a standard
training process that tunes a model to harmlessly follow user instructions—would
mitigate the risk of extraction. However, we develop two novel attacks that undo
a language model’s alignment and recover thousands of training examples from
popular proprietary aligned models such as OpenAI’s ChatGPT. Our work high-
lights the limitations of existing safeguards to prevent training data leakage in
production language models.
title: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
abstract: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
Speech language models (SpeechLMs) accept speech input and produce speech
output, allowing for more natural human-computer interaction compared to text-
based large language models (LLMs). Traditional approaches for developing
SpeechLMs are constrained by the limited availability of unsupervised speech
data and parallel speech-text data, which are significantly less abundant than text
pre-training data, thereby limiting their scalability as LLMs. We propose a novel
approach to scaling speech-text pre-training by leveraging large-scale synthetic in-
terleaved data derived from text corpora, eliminating the need for parallel speech-
text datasets. Our method efficiently constructs speech-text interleaved data by
sampling text spans from existing text corpora and synthesizing corresponding
speech spans using a text-to-token model, bypassing the need to generate actual
speech. We also employ a supervised speech tokenizer derived from an auto-
matic speech recognition (ASR) model by incorporating a vector-quantized bot-
tleneck into the encoder. This supervised training approach results in discrete
speech tokens with strong semantic preservation even at lower frame rates (e.g.
title: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
abstract: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
Synthetic data augmentation via Large Language Models (LLMs) allows re-
searchers to leverage additional training data, thus enhancing the performance of
downstream tasks, especially when real-world data is scarce. However, the gen-
erated data can deviate from the real-world data, and this misalignment can bring
about deficient results while applying the trained model to applications. There-
fore, we proposed efficient weighted-loss approaches to align synthetic data with
real-world distribution by emphasizing high-quality and diversified data generated
by LLMs using merely a tiny amount of real-world data. We empirically assessed
the effectiveness of our methods on multiple text classification tasks, and the re-
sults showed that leveraging our approaches on a BERT-level model robustly out-
performed standard cross-entropy and other data weighting approaches, providing
potential solutions to effectively leveraging synthetic data from any suitable data
generator.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Strong Model Collapse
abstract: Strong Model Collapse
Within the scaling laws paradigm, which underpins the training of large neural
networks like ChatGPT and Llama, we consider a supervised regression setting
and establish a strong form of the model collapse phenomenon, a critical perfor-
mance degradation due to synthetic data in the training corpus. Our results show
that even the smallest fraction of synthetic data (e.g., as little as 1 per 1000) can
still lead to model collapse: larger and larger training sets do not enhance perfor-
mance. We further investigate whether increasing model size, an approach aligned
with current trends in training large language models, exacerbates or mitigates
model collapse. In a simplified regime where neural networks are approximated
via random projections of tunable size, we both theoretically and empirically show
that larger models can amplify model collapse. Interestingly, our theory also in-
dicates that, beyond the interpolation threshold (which can be extremely high for
very large datasets), larger models may mitigate the collapse, although they do
not entirely prevent it. Our theoretical findings are empirically verified through
experiments on language models and neural networks for images.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
abstract: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
Training on high-quality synthetic data from strong language models (LMs) is a
common strategy to improve the reasoning performance of LMs. In this work,
we revisit whether this strategy is compute-optimal under a fixed inference bud-
get (e.g., FLOPs). To do so, we investigate the trade-offs between generating
synthetic data using a stronger but more expensive (SE) model versus a weaker
but cheaper (WC) model. We evaluate the generated data across three key met-
rics: coverage, diversity, and false positive rate, and show that the data from WC
models may have higher coverage and diversity, but also exhibit higher false pos-
itive rates. We then finetune LMs on data from SE and WC models in different
settings: knowledge distillation, self-improvement, and a novel weak-to-strong
improvement setup where a weaker LM teaches reasoning to a stronger LM. Our
findings reveal that models finetuned on WC-generated data consistently outper-
form those trained on SE-generated data across multiple benchmarks and multiple
choices of WC and SE models. These results challenge the prevailing practice of
relying on SE models for synthetic data generation, suggesting that WC may be
the compute-optimal approach for training advanced LM reasoners.
(a) Finetuning LMs with Gemma2 data.
(b) Finetuning LMs with Gemini 1.5 data.
Figure 1: Summary of the results. (a) We finetune Gemma-7B, Gemma2-9B, and Gemma2-27B on
the synthetic data collected from a stronger but more expensive LM (Gemma2-27B) and a weaker
but cheaper LM (Gemma2-9B) in a compute-matched setup for the MATH dataset. We find that
training with Gemma2-9B data is more compute-optimal across diverse finetuning paradigms –
knowledge distillation, self-improvement, and weak-to-strong improvement (i.e. using a weaker
model to improve a stronger model). (b) We finetune Gemma models (7B/9B/27B) on synthetic
data generated by Gemini-1.5-Pro and Gemini-1.5-Flash in a price-matched setup. We find that
finetuning with Flash-generated data consistently outperforms Pro-generated data.
∗Llama experiments in this paper were conducted only by parties outside of Google. Authors affiliated with
Google were responsible for the Gemini and Gemma experiments.
title: Self-Boosting Large Language Models with Synthetic Preference Data
abstract: Self-Boosting Large Language Models with Synthetic Preference Data
Through alignment with human preferences, Large Language Models (LLMs)
have advanced significantly in generating honest, harmless, and helpful responses.
However, collecting high-quality preference data is a resource-intensive and
creativity-demanding process, especially for the continual improvement of LLMs.
We introduce SynPO, a self-boosting paradigm that leverages synthetic prefer-
ence data for model alignment. SynPO employs an iterative mechanism wherein
a self-prompt generator creates diverse prompts, and a response improver refines
model responses progressively. This approach trains LLMs to autonomously learn
the generative rewards for their own outputs and eliminates the need for large-
scale annotation of prompts and human preferences. After four SynPO itera-
tions, Llama3-8B and Mistral-7B show significant enhancements in instruction-
following abilities, achieving over 22.1% win rate improvements on AlpacaEval
2.0 and ArenaHard. Simultaneously, SynPO improves the general performance
of LLMs on various tasks, validated by a 3.2 to 5.0 average score increase on the
well-recognized Open LLM leaderboard.
title: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
abstract: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
The rise of Large Language Models (LLMs) has accentuated the need for diverse,
high-quality pre-training data. Synthetic data emerges as a viable solution to the
challenges of data scarcity and inaccessibility. While previous literature has fo-
cused predominantly on the quality and quantity of real data, our work enables the
measurement of diversity in synthetic data and explores its impact on LLM perfor-
mance. We study the downstream effects of synthetic data diversity during both
the pre-training and fine-tuning stages by introducing a new diversity metric, LLM
cluster-agent, designed to evaluate the diversity of synthetic datasets. Through a
series of controlled experiments with models of 350M and 1.4B parameters, we
demonstrate that the proposed cluster-based LLM scoring of diversity correlates
positively with both pre-training and supervised fine-tuning performance. Our
findings also reveal that synthetic data diversity in pre-training affects supervised
fine-tuning more significantly than pre-training itself, even for smaller models. We
hope this study advances our understanding of the optimal use of synthetic data in
LLM training and opens new avenues for efficient data generation processes.
title: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
abstract: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
Speech language models (SpeechLMs) accept speech input and produce speech
output, allowing for more natural human-computer interaction compared to text-
based large language models (LLMs). Traditional approaches for developing
SpeechLMs are constrained by the limited availability of unsupervised speech
data and parallel speech-text data, which are significantly less abundant than text
pre-training data, thereby limiting their scalability as LLMs. We propose a novel
approach to scaling speech-text pre-training by leveraging large-scale synthetic in-
terleaved data derived from text corpora, eliminating the need for parallel speech-
text datasets. Our method efficiently constructs speech-text interleaved data by
sampling text spans from existing text corpora and synthesizing corresponding
speech spans using a text-to-token model, bypassing the need to generate actual
speech. We also employ a supervised speech tokenizer derived from an auto-
matic speech recognition (ASR) model by incorporating a vector-quantized bot-
tleneck into the encoder. This supervised training approach results in discrete
speech tokens with strong semantic preservation even at lower frame rates (e.g.
title: DataGen: Unified Synthetic Dataset Generation via Large Language Models
abstract: DataGen: Unified Synthetic Dataset Generation via Large Language Models
Large Language Models (LLMs) such as GPT-4 and Llama3 have significantly
impacted various fields by enabling high-quality synthetic data generation and
reducing dependence on expensive human-generated datasets. Despite this, chal-
lenges remain in the areas of generalization, controllability, diversity, and truthful-
ness within the existing generative frameworks. To address these challenges, this
paper presents DATAGEN, a comprehensive LLM-powered framework designed to
produce diverse, accurate, and highly controllable datasets. DATAGEN is adaptable,
supporting all types of text datasets and enhancing the generative process through
innovative mechanisms. To augment data diversity, DATAGEN incorporates an
attribute-guided generation module and a group checking feature. For accuracy, it
employs a code-based mathematical assessment for label verification alongside a
retrieval-augmented generation technique for factual validation. The framework
also allows for user-specified constraints, enabling customization of the data gener-
ation process to suit particular requirements. Extensive experiments demonstrate
the superior quality of data generated by DATAGEN, and each module within
DATAGEN plays a critical role in this enhancement. Additionally, DATAGEN is
applied in two practical scenarios: benchmarking LLMs and data augmentation.
The results indicate that DATAGEN effectively supports dynamic and evolving
benchmarking and that data augmentation improves LLM capabilities in various
domains, including agent-oriented abilities and reasoning skills.
title: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
abstract: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
Recent studies have shown that Large Language Models (LLMs) struggle to accu-
rately retrieve information and maintain reasoning capabilities when processing
long-context inputs. To address these limitations, we propose a finetuning approach
utilizing a carefully designed synthetic dataset comprising numerical key-value
retrieval tasks. Our experiments on models like GPT-3.5 Turbo and Mistral 7B
demonstrate that finetuning LLMs on this dataset significantly improves LLMs’ in-
formation retrieval and reasoning capabilities in longer-context settings. We present
an analysis of the finetuned models, illustrating the transfer of skills from synthetic
to real task evaluations (e.g., 10.5% improvement on 20 documents MDQA at
position 10 for GPT-3.5 Turbo). We also find that finetuned LLMs’ performance
on general benchmarks remains almost constant while LLMs finetuned on other
baseline long-context augmentation data can encourage hallucination (e.g., on
TriviaQA, Mistral 7B finetuned on our synthetic data cause no performance drop
while other baseline data can cause a drop that ranges from 2.33% to 6.19%). Our
study highlights the potential of finetuning on synthetic data for improving the
performance of LLMs on longer-context tasks.
title: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
abstract: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
We explore model collapse caused by synthetic data, where AI models trained on
such data experience a gradual decline in performance. Our initial analysis exam-
ines language model pretraining on mixed human and synthetic data, highlighting
performance degradation. Further statistical analysis reveals distributional shifts
and an over-concentration of n-gram features caused by synthetic data. Inspired
by these insights, we propose token-level editing on human data, to obtain semi-
synthetic data instead of fully using model outputs. As a proof of concept, we
theoretically demonstrate that token-level editing can prevent model collapse, as
the test error is constrained by a finite upper bound. We conducted extensive ex-
periments on pretraining, continual pretraining, and supervised fine-tuning of lan-
guage models. The results validate our theoretical proof that token-level editing
improves data quality and enhances model performance.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
abstract: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
Software engineers mainly write code by editing existing programs. In contrast,
language models (LMs) autoregressively synthesize programs in a single pass.
One explanation for this is the scarcity of sequential edit data. While high-quality
instruction data for code synthesis is scarce, edit data for synthesis is even scarcer.
To fill this gap, we develop a synthetic data generation algorithm called LintSeq.
This algorithm refactors programs into sequences of synthetic edits by using a
linter to procedurally sample across interdependent lines of source code. Synthetic
edits sampled with LintSeq reflect the syntax and semantics of their programming
language. To test the algorithm, we use it to refactor a dataset of instruction +
program pairs into instruction + program-diff-sequence tuples. Then, we fine-
tune a series of smaller LMs ranging from 2.6B to 14B parameters on both the
re-factored and original versions of this dataset. We perform comprehensive
evaluations comparing edit sequence code LMs against baselines on HumanEval,
MBPP(+), CodeContests, DS-1000, and BigCodeBench. We show that models
fine-tuned to iteratively synthesize code match or outperform baselines on pass@1,
and exhibit better scaling across higher pass@k as a function of total test-time
FLOPs. Finally, we also pretrain our own tiny LMs for code understanding. We
show that fine-tuning these models to synthesize code edit-by-edit results in strong
performance on HumanEval and MBPP(+) compared to existing code language
models of similar scale such as CodeT5+, AlphaCode, and Codex.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
abstract: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
Recent studies have shown that Large Language Models (LLMs) struggle to accu-
rately retrieve information and maintain reasoning capabilities when processing
long-context inputs. To address these limitations, we propose a finetuning approach
utilizing a carefully designed synthetic dataset comprising numerical key-value
retrieval tasks. Our experiments on models like GPT-3.5 Turbo and Mistral 7B
demonstrate that finetuning LLMs on this dataset significantly improves LLMs’ in-
formation retrieval and reasoning capabilities in longer-context settings. We present
an analysis of the finetuned models, illustrating the transfer of skills from synthetic
to real task evaluations (e.g., 10.5% improvement on 20 documents MDQA at
position 10 for GPT-3.5 Turbo). We also find that finetuned LLMs’ performance
on general benchmarks remains almost constant while LLMs finetuned on other
baseline long-context augmentation data can encourage hallucination (e.g., on
TriviaQA, Mistral 7B finetuned on our synthetic data cause no performance drop
while other baseline data can cause a drop that ranges from 2.33% to 6.19%). Our
study highlights the potential of finetuning on synthetic data for improving the
performance of LLMs on longer-context tasks.
title: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
abstract: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
We explore model collapse caused by synthetic data, where AI models trained on
such data experience a gradual decline in performance. Our initial analysis exam-
ines language model pretraining on mixed human and synthetic data, highlighting
performance degradation. Further statistical analysis reveals distributional shifts
and an over-concentration of n-gram features caused by synthetic data. Inspired
by these insights, we propose token-level editing on human data, to obtain semi-
synthetic data instead of fully using model outputs. As a proof of concept, we
theoretically demonstrate that token-level editing can prevent model collapse, as
the test error is constrained by a finite upper bound. We conducted extensive ex-
periments on pretraining, continual pretraining, and supervised fine-tuning of lan-
guage models. The results validate our theoretical proof that token-level editing
improves data quality and enhances model performance.
title: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
abstract: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
The utility of synthetic data to enhance pretraining data quality and hence to im-
prove downstream task accuracy has been widely explored in recent large lan-
guage models (LLMs). Yet, these approaches fall inadequate in complex, multi-
hop and mathematical reasoning tasks as the synthetic data typically fails to add
complementary knowledge to the existing raw corpus. In this work, we propose a
novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) gener-
ation method that improves the mathematical reasoning ability of LLMs. Specifi-
cally, using MIND, we generate synthetic conversations based on OpenWebMath
(OWM), resulting in a new math corpus, MIND-OWM. Our experiments with dif-
ferent conversational settings reveal that incorporating knowledge gaps between
dialog participants is essential for generating high-quality math data. We further
identify an effective way to format and integrate synthetic and raw data during pre-
training to maximize the gain in mathematical reasoning, emphasizing the need to
restructure raw data rather than use it as-is. Compared to pretraining just on raw
data, a model pretrained on MIND-OWM shows significant boost in mathematical
reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance
in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general
purpose reasoning tasks (GENERAL REASONING: +2.51%).
title: Synthetic continued pretraining
abstract: Synthetic continued pretraining
Pretraining on large-scale, unstructured internet text enables language models to
acquire a significant amount of world knowledge. However, this knowledge acqui-
sition is data-inefficient—to learn a fact, models must be trained on hundreds to
thousands of diverse representations of it. This poses a challenge when adapting a
pretrained model to a small corpus of domain-specific documents, where each fact
may appear rarely or only once. We propose to bridge this gap with synthetic con-
tinued pretraining: using the small domain-specific corpus to synthesize a large
corpus more amenable to learning, and then performing continued pretraining on
the synthesized corpus. We instantiate this proposal with EntiGraph, a synthetic
data augmentation algorithm that extracts salient entities from the source corpus
and then generates diverse text by drawing connections between those entities.
Synthetic continued pretraining with EntiGraph enables a language model to an-
swer questions and follow generic instructions related to the source documents
without access to them. If the source documents are instead available at inference
time, we show that the knowledge acquired through our approach compounds with
retrieval-augmented generation. To better understand these results, we build a sim-
ple mathematical model of EntiGraph, and show how synthetic data augmentation
can “rearrange” knowledge to enable more data-efficient learning.
title: Self-Boosting Large Language Models with Synthetic Preference Data
abstract: Self-Boosting Large Language Models with Synthetic Preference Data
Through alignment with human preferences, Large Language Models (LLMs)
have advanced significantly in generating honest, harmless, and helpful responses.
However, collecting high-quality preference data is a resource-intensive and
creativity-demanding process, especially for the continual improvement of LLMs.
We introduce SynPO, a self-boosting paradigm that leverages synthetic prefer-
ence data for model alignment. SynPO employs an iterative mechanism wherein
a self-prompt generator creates diverse prompts, and a response improver refines
model responses progressively. This approach trains LLMs to autonomously learn
the generative rewards for their own outputs and eliminates the need for large-
scale annotation of prompts and human preferences. After four SynPO itera-
tions, Llama3-8B and Mistral-7B show significant enhancements in instruction-
following abilities, achieving over 22.1% win rate improvements on AlpacaEval
2.0 and ArenaHard. Simultaneously, SynPO improves the general performance
of LLMs on various tasks, validated by a 3.2 to 5.0 average score increase on the
well-recognized Open LLM leaderboard.
title: Strong Model Collapse
abstract: Strong Model Collapse
Within the scaling laws paradigm, which underpins the training of large neural
networks like ChatGPT and Llama, we consider a supervised regression setting
and establish a strong form of the model collapse phenomenon, a critical perfor-
mance degradation due to synthetic data in the training corpus. Our results show
that even the smallest fraction of synthetic data (e.g., as little as 1 per 1000) can
still lead to model collapse: larger and larger training sets do not enhance perfor-
mance. We further investigate whether increasing model size, an approach aligned
with current trends in training large language models, exacerbates or mitigates
model collapse. In a simplified regime where neural networks are approximated
via random projections of tunable size, we both theoretically and empirically show
that larger models can amplify model collapse. Interestingly, our theory also in-
dicates that, beyond the interpolation threshold (which can be extremely high for
very large datasets), larger models may mitigate the collapse, although they do
not entirely prevent it. Our theoretical findings are empirically verified through
experiments on language models and neural networks for images.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: Synthetic continued pretraining
abstract: Synthetic continued pretraining
Pretraining on large-scale, unstructured internet text enables language models to
acquire a significant amount of world knowledge. However, this knowledge acqui-
sition is data-inefficient—to learn a fact, models must be trained on hundreds to
thousands of diverse representations of it. This poses a challenge when adapting a
pretrained model to a small corpus of domain-specific documents, where each fact
may appear rarely or only once. We propose to bridge this gap with synthetic con-
tinued pretraining: using the small domain-specific corpus to synthesize a large
corpus more amenable to learning, and then performing continued pretraining on
the synthesized corpus. We instantiate this proposal with EntiGraph, a synthetic
data augmentation algorithm that extracts salient entities from the source corpus
and then generates diverse text by drawing connections between those entities.
Synthetic continued pretraining with EntiGraph enables a language model to an-
swer questions and follow generic instructions related to the source documents
without access to them. If the source documents are instead available at inference
time, we show that the knowledge acquired through our approach compounds with
retrieval-augmented generation. To better understand these results, we build a sim-
ple mathematical model of EntiGraph, and show how synthetic data augmentation
can “rearrange” knowledge to enable more data-efficient learning.
title: DataGen: Unified Synthetic Dataset Generation via Large Language Models
abstract: DataGen: Unified Synthetic Dataset Generation via Large Language Models
Large Language Models (LLMs) such as GPT-4 and Llama3 have significantly
impacted various fields by enabling high-quality synthetic data generation and
reducing dependence on expensive human-generated datasets. Despite this, chal-
lenges remain in the areas of generalization, controllability, diversity, and truthful-
ness within the existing generative frameworks. To address these challenges, this
paper presents DATAGEN, a comprehensive LLM-powered framework designed to
produce diverse, accurate, and highly controllable datasets. DATAGEN is adaptable,
supporting all types of text datasets and enhancing the generative process through
innovative mechanisms. To augment data diversity, DATAGEN incorporates an
attribute-guided generation module and a group checking feature. For accuracy, it
employs a code-based mathematical assessment for label verification alongside a
retrieval-augmented generation technique for factual validation. The framework
also allows for user-specified constraints, enabling customization of the data gener-
ation process to suit particular requirements. Extensive experiments demonstrate
the superior quality of data generated by DATAGEN, and each module within
DATAGEN plays a critical role in this enhancement. Additionally, DATAGEN is
applied in two practical scenarios: benchmarking LLMs and data augmentation.
The results indicate that DATAGEN effectively supports dynamic and evolving
benchmarking and that data augmentation improves LLM capabilities in various
domains, including agent-oriented abilities and reasoning skills.
title: Scalable Extraction of Training Data from Aligned, Production Language Models
abstract: Scalable Extraction of Training Data from Aligned, Production Language Models
Large language models are prone to memorizing some of their training data. Mem-
orized (and possibly sensitive) samples can then be extracted at generation time
by adversarial or benign users. There is hope that model alignment—a standard
training process that tunes a model to harmlessly follow user instructions—would
mitigate the risk of extraction. However, we develop two novel attacks that undo
a language model’s alignment and recover thousands of training examples from
popular proprietary aligned models such as OpenAI’s ChatGPT. Our work high-
lights the limitations of existing safeguards to prevent training data leakage in
production language models.
title: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
abstract: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
We explore model collapse caused by synthetic data, where AI models trained on
such data experience a gradual decline in performance. Our initial analysis exam-
ines language model pretraining on mixed human and synthetic data, highlighting
performance degradation. Further statistical analysis reveals distributional shifts
and an over-concentration of n-gram features caused by synthetic data. Inspired
by these insights, we propose token-level editing on human data, to obtain semi-
synthetic data instead of fully using model outputs. As a proof of concept, we
theoretically demonstrate that token-level editing can prevent model collapse, as
the test error is constrained by a finite upper bound. We conducted extensive ex-
periments on pretraining, continual pretraining, and supervised fine-tuning of lan-
guage models. The results validate our theoretical proof that token-level editing
improves data quality and enhances model performance.
title: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
abstract: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
Biological language model performance depends heavily on pretraining data
quality, diversity, and size. While metagenomic datasets feature enor-
mous biological diversity, their utilization as pretraining data has been
limited due to challenges in data accessibility, quality filtering and dedupli-
cation. Here, we present the Open MetaGenomic (OMG) corpus, a genomic
pretraining dataset totalling 3.1T base pairs and 3.3B protein coding se-
quences, obtained by combining two largest metagenomic dataset reposito-
ries (JGI’s IMG and EMBL’s MGnify). We first document the composition
of the dataset and describe the quality filtering steps taken to remove poor
quality data. We make the OMG corpus available as a mixed-modality
genomic sequence dataset that represents multi-gene encoding genomic
sequences with translated amino acids for protein coding sequences, and
nucleic acids for intergenic sequences. We train the first mixed-modality
genomic language model (gLM2) that leverages genomic context informa-
tion to learn robust functional representations, as well as coevolutionary
signals in protein-protein interfaces and genomic regulatory syntax. Fur-
thermore, we show that deduplication in embedding space can be used to
balance the corpus, demonstrating improved performance on downstream
tasks. The OMG dataset is publicly hosted on the Hugging Face Hub
at https://huggingface.co/datasets/tattabio/OMG and gLM2 is avail-
able at https://huggingface.co/tattabio/gLM2_650M.
title: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
abstract: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
Training on high-quality synthetic data from strong language models (LMs) is a
common strategy to improve the reasoning performance of LMs. In this work,
we revisit whether this strategy is compute-optimal under a fixed inference bud-
get (e.g., FLOPs). To do so, we investigate the trade-offs between generating
synthetic data using a stronger but more expensive (SE) model versus a weaker
but cheaper (WC) model. We evaluate the generated data across three key met-
rics: coverage, diversity, and false positive rate, and show that the data from WC
models may have higher coverage and diversity, but also exhibit higher false pos-
itive rates. We then finetune LMs on data from SE and WC models in different
settings: knowledge distillation, self-improvement, and a novel weak-to-strong
improvement setup where a weaker LM teaches reasoning to a stronger LM. Our
findings reveal that models finetuned on WC-generated data consistently outper-
form those trained on SE-generated data across multiple benchmarks and multiple
choices of WC and SE models. These results challenge the prevailing practice of
relying on SE models for synthetic data generation, suggesting that WC may be
the compute-optimal approach for training advanced LM reasoners.
(a) Finetuning LMs with Gemma2 data.
(b) Finetuning LMs with Gemini 1.5 data.
Figure 1: Summary of the results. (a) We finetune Gemma-7B, Gemma2-9B, and Gemma2-27B on
the synthetic data collected from a stronger but more expensive LM (Gemma2-27B) and a weaker
but cheaper LM (Gemma2-9B) in a compute-matched setup for the MATH dataset. We find that
training with Gemma2-9B data is more compute-optimal across diverse finetuning paradigms –
knowledge distillation, self-improvement, and weak-to-strong improvement (i.e. using a weaker
model to improve a stronger model). (b) We finetune Gemma models (7B/9B/27B) on synthetic
data generated by Gemini-1.5-Pro and Gemini-1.5-Flash in a price-matched setup. We find that
finetuning with Flash-generated data consistently outperforms Pro-generated data.
∗Llama experiments in this paper were conducted only by parties outside of Google. Authors affiliated with
Google were responsible for the Gemini and Gemma experiments.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
abstract: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
Recent studies have shown that Large Language Models (LLMs) struggle to accu-
rately retrieve information and maintain reasoning capabilities when processing
long-context inputs. To address these limitations, we propose a finetuning approach
utilizing a carefully designed synthetic dataset comprising numerical key-value
retrieval tasks. Our experiments on models like GPT-3.5 Turbo and Mistral 7B
demonstrate that finetuning LLMs on this dataset significantly improves LLMs’ in-
formation retrieval and reasoning capabilities in longer-context settings. We present
an analysis of the finetuned models, illustrating the transfer of skills from synthetic
to real task evaluations (e.g., 10.5% improvement on 20 documents MDQA at
position 10 for GPT-3.5 Turbo). We also find that finetuned LLMs’ performance
on general benchmarks remains almost constant while LLMs finetuned on other
baseline long-context augmentation data can encourage hallucination (e.g., on
TriviaQA, Mistral 7B finetuned on our synthetic data cause no performance drop
while other baseline data can cause a drop that ranges from 2.33% to 6.19%). Our
study highlights the potential of finetuning on synthetic data for improving the
performance of LLMs on longer-context tasks.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
abstract: MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
The utility of synthetic data to enhance pretraining data quality and hence to im-
prove downstream task accuracy has been widely explored in recent large lan-
guage models (LLMs). Yet, these approaches fall inadequate in complex, multi-
hop and mathematical reasoning tasks as the synthetic data typically fails to add
complementary knowledge to the existing raw corpus. In this work, we propose a
novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) gener-
ation method that improves the mathematical reasoning ability of LLMs. Specifi-
cally, using MIND, we generate synthetic conversations based on OpenWebMath
(OWM), resulting in a new math corpus, MIND-OWM. Our experiments with dif-
ferent conversational settings reveal that incorporating knowledge gaps between
dialog participants is essential for generating high-quality math data. We further
identify an effective way to format and integrate synthetic and raw data during pre-
training to maximize the gain in mathematical reasoning, emphasizing the need to
restructure raw data rather than use it as-is. Compared to pretraining just on raw
data, a model pretrained on MIND-OWM shows significant boost in mathematical
reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance
in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general
purpose reasoning tasks (GENERAL REASONING: +2.51%).
title: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
abstract: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
Biological language model performance depends heavily on pretraining data
quality, diversity, and size. While metagenomic datasets feature enor-
mous biological diversity, their utilization as pretraining data has been
limited due to challenges in data accessibility, quality filtering and dedupli-
cation. Here, we present the Open MetaGenomic (OMG) corpus, a genomic
pretraining dataset totalling 3.1T base pairs and 3.3B protein coding se-
quences, obtained by combining two largest metagenomic dataset reposito-
ries (JGI’s IMG and EMBL’s MGnify). We first document the composition
of the dataset and describe the quality filtering steps taken to remove poor
quality data. We make the OMG corpus available as a mixed-modality
genomic sequence dataset that represents multi-gene encoding genomic
sequences with translated amino acids for protein coding sequences, and
nucleic acids for intergenic sequences. We train the first mixed-modality
genomic language model (gLM2) that leverages genomic context informa-
tion to learn robust functional representations, as well as coevolutionary
signals in protein-protein interfaces and genomic regulatory syntax. Fur-
thermore, we show that deduplication in embedding space can be used to
balance the corpus, demonstrating improved performance on downstream
tasks. The OMG dataset is publicly hosted on the Hugging Face Hub
at https://huggingface.co/datasets/tattabio/OMG and gLM2 is avail-
able at https://huggingface.co/tattabio/gLM2_650M.
title: Scalable Extraction of Training Data from Aligned, Production Language Models
abstract: Scalable Extraction of Training Data from Aligned, Production Language Models
Large language models are prone to memorizing some of their training data. Mem-
orized (and possibly sensitive) samples can then be extracted at generation time
by adversarial or benign users. There is hope that model alignment—a standard
training process that tunes a model to harmlessly follow user instructions—would
mitigate the risk of extraction. However, we develop two novel attacks that undo
a language model’s alignment and recover thousands of training examples from
popular proprietary aligned models such as OpenAI’s ChatGPT. Our work high-
lights the limitations of existing safeguards to prevent training data leakage in
production language models.
title: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
abstract: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
Synthetic data has been widely used to train large language models, but their gener-
ative nature inevitably introduces noisy, non-informative, and misleading learning
signals. In this paper, we propose MONTESSORI-INSTRUCT, a novel data synthesis
framework that tailors the data synthesis ability of the teacher language model
toward the student language model’s learning process. Specifically, we utilize local
data influence of synthetic training data points on students to characterize students’
learning preferences. Then, we train the teacher model with Direct Preference
Optimization (DPO) to generate synthetic data tailored toward student learning pref-
erences. Experiments with Llama3-8B-Instruct (teacher) and Llama3-8B (student)
on Alpaca Eval and MT-Bench demonstrate that Montessori-Instruct significantly
outperforms standard synthesis methods by 18.35% and 46.24% relatively. Our
method also beats data synthesized by a stronger teacher model, GPT-4o. Further
analysis confirms the benefits of teacher’s learning to generate more influential train-
ing data in the student’s improved learning, the advantages of local data influence
in accurately measuring student preferences, and the robustness of Montessori-
Instruct across different student models. Our code and data are open-sourced at
https://github.com/cxcscmu/Montessori-Instruct.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective
abstract: Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective
Synthetic data has become a pivotal resource in post-training tasks for large lan-
guage models (LLMs) due to the scarcity of high-quality, specific data. While
various methods have been developed to generate synthetic data, there remains a
discernible gap between the practical effects of synthetic data and our theoretical
comprehension. To address this challenge, we commence by presenting a detailed
modeling of the prevalent synthetic data generation process. Building upon this
modeling, we demonstrate that the generalization capability of the post-trained
model is critically determined by the information gain derived from the generative
model, as analyzed from a novel reverse-bottleneck perspective. Moreover, we in-
troduce the concept of Generalization Gain via Mutual Information (GGMI) and
elucidate the relationship between generalization gain and information gain. This
analysis serves as a theoretical foundation for synthetic data generation and further
highlights its connection with the generalization capability of post-trained models,
offering an understanding about the design of synthetic data generation techniques
and the optimization of the post-training process. We open-source our code at
https://github.com/ZyGan1999/Towards-a-Theoretical-U
nderstanding-of-Synthetic-Data-in-LLM-Post-Training.
title: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
abstract: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
The rise of Large Language Models (LLMs) has accentuated the need for diverse,
high-quality pre-training data. Synthetic data emerges as a viable solution to the
challenges of data scarcity and inaccessibility. While previous literature has fo-
cused predominantly on the quality and quantity of real data, our work enables the
measurement of diversity in synthetic data and explores its impact on LLM perfor-
mance. We study the downstream effects of synthetic data diversity during both
the pre-training and fine-tuning stages by introducing a new diversity metric, LLM
cluster-agent, designed to evaluate the diversity of synthetic datasets. Through a
series of controlled experiments with models of 350M and 1.4B parameters, we
demonstrate that the proposed cluster-based LLM scoring of diversity correlates
positively with both pre-training and supervised fine-tuning performance. Our
findings also reveal that synthetic data diversity in pre-training affects supervised
fine-tuning more significantly than pre-training itself, even for smaller models. We
hope this study advances our understanding of the optimal use of synthetic data in
LLM training and opens new avenues for efficient data generation processes.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
abstract: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
Synthetic data augmentation via Large Language Models (LLMs) allows re-
searchers to leverage additional training data, thus enhancing the performance of
downstream tasks, especially when real-world data is scarce. However, the gen-
erated data can deviate from the real-world data, and this misalignment can bring
about deficient results while applying the trained model to applications. There-
fore, we proposed efficient weighted-loss approaches to align synthetic data with
real-world distribution by emphasizing high-quality and diversified data generated
by LLMs using merely a tiny amount of real-world data. We empirically assessed
the effectiveness of our methods on multiple text classification tasks, and the re-
sults showed that leveraging our approaches on a BERT-level model robustly out-
performed standard cross-entropy and other data weighting approaches, providing
potential solutions to effectively leveraging synthetic data from any suitable data
generator.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
abstract: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
Recent studies have shown that Large Language Models (LLMs) struggle to accu-
rately retrieve information and maintain reasoning capabilities when processing
long-context inputs. To address these limitations, we propose a finetuning approach
utilizing a carefully designed synthetic dataset comprising numerical key-value
retrieval tasks. Our experiments on models like GPT-3.5 Turbo and Mistral 7B
demonstrate that finetuning LLMs on this dataset significantly improves LLMs’ in-
formation retrieval and reasoning capabilities in longer-context settings. We present
an analysis of the finetuned models, illustrating the transfer of skills from synthetic
to real task evaluations (e.g., 10.5% improvement on 20 documents MDQA at
position 10 for GPT-3.5 Turbo). We also find that finetuned LLMs’ performance
on general benchmarks remains almost constant while LLMs finetuned on other
baseline long-context augmentation data can encourage hallucination (e.g., on
TriviaQA, Mistral 7B finetuned on our synthetic data cause no performance drop
while other baseline data can cause a drop that ranges from 2.33% to 6.19%). Our
study highlights the potential of finetuning on synthetic data for improving the
performance of LLMs on longer-context tasks.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: Strong Model Collapse
abstract: Strong Model Collapse
Within the scaling laws paradigm, which underpins the training of large neural
networks like ChatGPT and Llama, we consider a supervised regression setting
and establish a strong form of the model collapse phenomenon, a critical perfor-
mance degradation due to synthetic data in the training corpus. Our results show
that even the smallest fraction of synthetic data (e.g., as little as 1 per 1000) can
still lead to model collapse: larger and larger training sets do not enhance perfor-
mance. We further investigate whether increasing model size, an approach aligned
with current trends in training large language models, exacerbates or mitigates
model collapse. In a simplified regime where neural networks are approximated
via random projections of tunable size, we both theoretically and empirically show
that larger models can amplify model collapse. Interestingly, our theory also in-
dicates that, beyond the interpolation threshold (which can be extremely high for
very large datasets), larger models may mitigate the collapse, although they do
not entirely prevent it. Our theoretical findings are empirically verified through
experiments on language models and neural networks for images.
title: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
abstract: The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
Biological language model performance depends heavily on pretraining data
quality, diversity, and size. While metagenomic datasets feature enor-
mous biological diversity, their utilization as pretraining data has been
limited due to challenges in data accessibility, quality filtering and dedupli-
cation. Here, we present the Open MetaGenomic (OMG) corpus, a genomic
pretraining dataset totalling 3.1T base pairs and 3.3B protein coding se-
quences, obtained by combining two largest metagenomic dataset reposito-
ries (JGI’s IMG and EMBL’s MGnify). We first document the composition
of the dataset and describe the quality filtering steps taken to remove poor
quality data. We make the OMG corpus available as a mixed-modality
genomic sequence dataset that represents multi-gene encoding genomic
sequences with translated amino acids for protein coding sequences, and
nucleic acids for intergenic sequences. We train the first mixed-modality
genomic language model (gLM2) that leverages genomic context informa-
tion to learn robust functional representations, as well as coevolutionary
signals in protein-protein interfaces and genomic regulatory syntax. Fur-
thermore, we show that deduplication in embedding space can be used to
balance the corpus, demonstrating improved performance on downstream
tasks. The OMG dataset is publicly hosted on the Hugging Face Hub
at https://huggingface.co/datasets/tattabio/OMG and gLM2 is avail-
able at https://huggingface.co/tattabio/gLM2_650M.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
abstract: ToEdit: How to Synthesize Text Data to Avoid Model Collapse?
We explore model collapse caused by synthetic data, where AI models trained on
such data experience a gradual decline in performance. Our initial analysis exam-
ines language model pretraining on mixed human and synthetic data, highlighting
performance degradation. Further statistical analysis reveals distributional shifts
and an over-concentration of n-gram features caused by synthetic data. Inspired
by these insights, we propose token-level editing on human data, to obtain semi-
synthetic data instead of fully using model outputs. As a proof of concept, we
theoretically demonstrate that token-level editing can prevent model collapse, as
the test error is constrained by a finite upper bound. We conducted extensive ex-
periments on pretraining, continual pretraining, and supervised fine-tuning of lan-
guage models. The results validate our theoretical proof that token-level editing
improves data quality and enhances model performance.
title: Synthetic continued pretraining
abstract: Synthetic continued pretraining
Pretraining on large-scale, unstructured internet text enables language models to
acquire a significant amount of world knowledge. However, this knowledge acqui-
sition is data-inefficient—to learn a fact, models must be trained on hundreds to
thousands of diverse representations of it. This poses a challenge when adapting a
pretrained model to a small corpus of domain-specific documents, where each fact
may appear rarely or only once. We propose to bridge this gap with synthetic con-
tinued pretraining: using the small domain-specific corpus to synthesize a large
corpus more amenable to learning, and then performing continued pretraining on
the synthesized corpus. We instantiate this proposal with EntiGraph, a synthetic
data augmentation algorithm that extracts salient entities from the source corpus
and then generates diverse text by drawing connections between those entities.
Synthetic continued pretraining with EntiGraph enables a language model to an-
swer questions and follow generic instructions related to the source documents
without access to them. If the source documents are instead available at inference
time, we show that the knowledge acquired through our approach compounds with
retrieval-augmented generation. To better understand these results, we build a sim-
ple mathematical model of EntiGraph, and show how synthetic data augmentation
can “rearrange” knowledge to enable more data-efficient learning.
title: Scalable Extraction of Training Data from Aligned, Production Language Models
abstract: Scalable Extraction of Training Data from Aligned, Production Language Models
Large language models are prone to memorizing some of their training data. Mem-
orized (and possibly sensitive) samples can then be extracted at generation time
by adversarial or benign users. There is hope that model alignment—a standard
training process that tunes a model to harmlessly follow user instructions—would
mitigate the risk of extraction. However, we develop two novel attacks that undo
a language model’s alignment and recover thousands of training examples from
popular proprietary aligned models such as OpenAI’s ChatGPT. Our work high-
lights the limitations of existing safeguards to prevent training data leakage in
production language models.
title: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
abstract: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
Software engineers mainly write code by editing existing programs. In contrast,
language models (LMs) autoregressively synthesize programs in a single pass.
One explanation for this is the scarcity of sequential edit data. While high-quality
instruction data for code synthesis is scarce, edit data for synthesis is even scarcer.
To fill this gap, we develop a synthetic data generation algorithm called LintSeq.
This algorithm refactors programs into sequences of synthetic edits by using a
linter to procedurally sample across interdependent lines of source code. Synthetic
edits sampled with LintSeq reflect the syntax and semantics of their programming
language. To test the algorithm, we use it to refactor a dataset of instruction +
program pairs into instruction + program-diff-sequence tuples. Then, we fine-
tune a series of smaller LMs ranging from 2.6B to 14B parameters on both the
re-factored and original versions of this dataset. We perform comprehensive
evaluations comparing edit sequence code LMs against baselines on HumanEval,
MBPP(+), CodeContests, DS-1000, and BigCodeBench. We show that models
fine-tuned to iteratively synthesize code match or outperform baselines on pass@1,
and exhibit better scaling across higher pass@k as a function of total test-time
FLOPs. Finally, we also pretrain our own tiny LMs for code understanding. We
show that fine-tuning these models to synthesize code edit-by-edit results in strong
performance on HumanEval and MBPP(+) compared to existing code language
models of similar scale such as CodeT5+, AlphaCode, and Codex.
title: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
abstract: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
Synthetic data has been widely used to train large language models, but their gener-
ative nature inevitably introduces noisy, non-informative, and misleading learning
signals. In this paper, we propose MONTESSORI-INSTRUCT, a novel data synthesis
framework that tailors the data synthesis ability of the teacher language model
toward the student language model’s learning process. Specifically, we utilize local
data influence of synthetic training data points on students to characterize students’
learning preferences. Then, we train the teacher model with Direct Preference
Optimization (DPO) to generate synthetic data tailored toward student learning pref-
erences. Experiments with Llama3-8B-Instruct (teacher) and Llama3-8B (student)
on Alpaca Eval and MT-Bench demonstrate that Montessori-Instruct significantly
outperforms standard synthesis methods by 18.35% and 46.24% relatively. Our
method also beats data synthesized by a stronger teacher model, GPT-4o. Further
analysis confirms the benefits of teacher’s learning to generate more influential train-
ing data in the student’s improved learning, the advantages of local data influence
in accurately measuring student preferences, and the robustness of Montessori-
Instruct across different student models. Our code and data are open-sourced at
https://github.com/cxcscmu/Montessori-Instruct.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
synthetic_data | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: synthetic data for training large language models.
Here are some relevant papers on this topic just for your background knowledge:
title: Strong Model Collapse
abstract: Strong Model Collapse
Within the scaling laws paradigm, which underpins the training of large neural
networks like ChatGPT and Llama, we consider a supervised regression setting
and establish a strong form of the model collapse phenomenon, a critical perfor-
mance degradation due to synthetic data in the training corpus. Our results show
that even the smallest fraction of synthetic data (e.g., as little as 1 per 1000) can
still lead to model collapse: larger and larger training sets do not enhance perfor-
mance. We further investigate whether increasing model size, an approach aligned
with current trends in training large language models, exacerbates or mitigates
model collapse. In a simplified regime where neural networks are approximated
via random projections of tunable size, we both theoretically and empirically show
that larger models can amplify model collapse. Interestingly, our theory also in-
dicates that, beyond the interpolation threshold (which can be extremely high for
very large datasets), larger models may mitigate the collapse, although they do
not entirely prevent it. Our theoretical findings are empirically verified through
experiments on language models and neural networks for images.
title: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
abstract: Scaling Speech-Text Pre-training with Synthetic Interleaved Data
Speech language models (SpeechLMs) accept speech input and produce speech
output, allowing for more natural human-computer interaction compared to text-
based large language models (LLMs). Traditional approaches for developing
SpeechLMs are constrained by the limited availability of unsupervised speech
data and parallel speech-text data, which are significantly less abundant than text
pre-training data, thereby limiting their scalability as LLMs. We propose a novel
approach to scaling speech-text pre-training by leveraging large-scale synthetic in-
terleaved data derived from text corpora, eliminating the need for parallel speech-
text datasets. Our method efficiently constructs speech-text interleaved data by
sampling text spans from existing text corpora and synthesizing corresponding
speech spans using a text-to-token model, bypassing the need to generate actual
speech. We also employ a supervised speech tokenizer derived from an auto-
matic speech recognition (ASR) model by incorporating a vector-quantized bot-
tleneck into the encoder. This supervised training approach results in discrete
speech tokens with strong semantic preservation even at lower frame rates (e.g.
title: DataMan: Data Manager for Pre-training Large Language Models
abstract: DataMan: Data Manager for Pre-training Large Language Models
The performance emergence of large language models (LLMs) driven by data
scaling laws makes the selection of pre-training data increasingly important. How-
ever, existing methods rely on limited heuristics and human intuition, lacking
comprehensive and clear guidelines. To address this, we are inspired by “reverse
thinking” – prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality
criteria from the causes of text perplexity anomalies and introduce 15 common
application domains to support domain mixing. In this paper, we train a Data
Manager (DataMan) to learn quality ratings and domain recognition from point-
wise rating, and use it to annotate a 447B token pre-training corpus with 14 quality
ratings and domain type. Our experiments validate our approach, using DataMan
to select 30B tokens to train a 1.3B-parameter language model, demonstrating
significant improvements in in-context learning (ICL), perplexity, and instruction-
following ability over the state-of-the-art baseline. The best-performing model,
based on the Overall Score l=5 surpasses a model trained with 50% more data
using uniform sampling. We continue pre-training with high-rated, domain-specific
data annotated by DataMan to enhance domain-specific ICL performance and thus
verify DataMan’s domain mixing ability. Our findings emphasize the importance of
quality ranking, the complementary nature of quality criteria, and their low correla-
tion with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition,
the distribution of quality ratings, and the original document sources.
title: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
abstract: From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
Recent studies have shown that Large Language Models (LLMs) struggle to accu-
rately retrieve information and maintain reasoning capabilities when processing
long-context inputs. To address these limitations, we propose a finetuning approach
utilizing a carefully designed synthetic dataset comprising numerical key-value
retrieval tasks. Our experiments on models like GPT-3.5 Turbo and Mistral 7B
demonstrate that finetuning LLMs on this dataset significantly improves LLMs’ in-
formation retrieval and reasoning capabilities in longer-context settings. We present
an analysis of the finetuned models, illustrating the transfer of skills from synthetic
to real task evaluations (e.g., 10.5% improvement on 20 documents MDQA at
position 10 for GPT-3.5 Turbo). We also find that finetuned LLMs’ performance
on general benchmarks remains almost constant while LLMs finetuned on other
baseline long-context augmentation data can encourage hallucination (e.g., on
TriviaQA, Mistral 7B finetuned on our synthetic data cause no performance drop
while other baseline data can cause a drop that ranges from 2.33% to 6.19%). Our
study highlights the potential of finetuning on synthetic data for improving the
performance of LLMs on longer-context tasks.
title: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
abstract: Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
Synthetic data augmentation via Large Language Models (LLMs) allows re-
searchers to leverage additional training data, thus enhancing the performance of
downstream tasks, especially when real-world data is scarce. However, the gen-
erated data can deviate from the real-world data, and this misalignment can bring
about deficient results while applying the trained model to applications. There-
fore, we proposed efficient weighted-loss approaches to align synthetic data with
real-world distribution by emphasizing high-quality and diversified data generated
by LLMs using merely a tiny amount of real-world data. We empirically assessed
the effectiveness of our methods on multiple text classification tasks, and the re-
sults showed that leveraging our approaches on a BERT-level model robustly out-
performed standard cross-entropy and other data weighting approaches, providing
potential solutions to effectively leveraging synthetic data from any suitable data
generator.
title: DataGen: Unified Synthetic Dataset Generation via Large Language Models
abstract: DataGen: Unified Synthetic Dataset Generation via Large Language Models
Large Language Models (LLMs) such as GPT-4 and Llama3 have significantly
impacted various fields by enabling high-quality synthetic data generation and
reducing dependence on expensive human-generated datasets. Despite this, chal-
lenges remain in the areas of generalization, controllability, diversity, and truthful-
ness within the existing generative frameworks. To address these challenges, this
paper presents DATAGEN, a comprehensive LLM-powered framework designed to
produce diverse, accurate, and highly controllable datasets. DATAGEN is adaptable,
supporting all types of text datasets and enhancing the generative process through
innovative mechanisms. To augment data diversity, DATAGEN incorporates an
attribute-guided generation module and a group checking feature. For accuracy, it
employs a code-based mathematical assessment for label verification alongside a
retrieval-augmented generation technique for factual validation. The framework
also allows for user-specified constraints, enabling customization of the data gener-
ation process to suit particular requirements. Extensive experiments demonstrate
the superior quality of data generated by DATAGEN, and each module within
DATAGEN plays a critical role in this enhancement. Additionally, DATAGEN is
applied in two practical scenarios: benchmarking LLMs and data augmentation.
The results indicate that DATAGEN effectively supports dynamic and evolving
benchmarking and that data augmentation improves LLM capabilities in various
domains, including agent-oriented abilities and reasoning skills.
title: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
abstract: Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
Synthetic data has been widely used to train large language models, but their gener-
ative nature inevitably introduces noisy, non-informative, and misleading learning
signals. In this paper, we propose MONTESSORI-INSTRUCT, a novel data synthesis
framework that tailors the data synthesis ability of the teacher language model
toward the student language model’s learning process. Specifically, we utilize local
data influence of synthetic training data points on students to characterize students’
learning preferences. Then, we train the teacher model with Direct Preference
Optimization (DPO) to generate synthetic data tailored toward student learning pref-
erences. Experiments with Llama3-8B-Instruct (teacher) and Llama3-8B (student)
on Alpaca Eval and MT-Bench demonstrate that Montessori-Instruct significantly
outperforms standard synthesis methods by 18.35% and 46.24% relatively. Our
method also beats data synthesized by a stronger teacher model, GPT-4o. Further
analysis confirms the benefits of teacher’s learning to generate more influential train-
ing data in the student’s improved learning, the advantages of local data influence
in accurately measuring student preferences, and the robustness of Montessori-
Instruct across different student models. Our code and data are open-sourced at
https://github.com/cxcscmu/Montessori-Instruct.
title: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
abstract: On the Diversity of Synthetic Data and its Impact on Training Large Language Models
The rise of Large Language Models (LLMs) has accentuated the need for diverse,
high-quality pre-training data. Synthetic data emerges as a viable solution to the
challenges of data scarcity and inaccessibility. While previous literature has fo-
cused predominantly on the quality and quantity of real data, our work enables the
measurement of diversity in synthetic data and explores its impact on LLM perfor-
mance. We study the downstream effects of synthetic data diversity during both
the pre-training and fine-tuning stages by introducing a new diversity metric, LLM
cluster-agent, designed to evaluate the diversity of synthetic datasets. Through a
series of controlled experiments with models of 350M and 1.4B parameters, we
demonstrate that the proposed cluster-based LLM scoring of diversity correlates
positively with both pre-training and supervised fine-tuning performance. Our
findings also reveal that synthetic data diversity in pre-training affects supervised
fine-tuning more significantly than pre-training itself, even for smaller models. We
hope this study advances our understanding of the optimal use of synthetic data in
LLM training and opens new avenues for efficient data generation processes.
title: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
abstract: Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
Software engineers mainly write code by editing existing programs. In contrast,
language models (LMs) autoregressively synthesize programs in a single pass.
One explanation for this is the scarcity of sequential edit data. While high-quality
instruction data for code synthesis is scarce, edit data for synthesis is even scarcer.
To fill this gap, we develop a synthetic data generation algorithm called LintSeq.
This algorithm refactors programs into sequences of synthetic edits by using a
linter to procedurally sample across interdependent lines of source code. Synthetic
edits sampled with LintSeq reflect the syntax and semantics of their programming
language. To test the algorithm, we use it to refactor a dataset of instruction +
program pairs into instruction + program-diff-sequence tuples. Then, we fine-
tune a series of smaller LMs ranging from 2.6B to 14B parameters on both the
re-factored and original versions of this dataset. We perform comprehensive
evaluations comparing edit sequence code LMs against baselines on HumanEval,
MBPP(+), CodeContests, DS-1000, and BigCodeBench. We show that models
fine-tuned to iteratively synthesize code match or outperform baselines on pass@1,
and exhibit better scaling across higher pass@k as a function of total test-time
FLOPs. Finally, we also pretrain our own tiny LMs for code understanding. We
show that fine-tuning these models to synthesize code edit-by-edit results in strong
performance on HumanEval and MBPP(+) compared to existing code language
models of similar scale such as CodeT5+, AlphaCode, and Codex.
title: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
abstract: Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
Training on high-quality synthetic data from strong language models (LMs) is a
common strategy to improve the reasoning performance of LMs. In this work,
we revisit whether this strategy is compute-optimal under a fixed inference bud-
get (e.g., FLOPs). To do so, we investigate the trade-offs between generating
synthetic data using a stronger but more expensive (SE) model versus a weaker
but cheaper (WC) model. We evaluate the generated data across three key met-
rics: coverage, diversity, and false positive rate, and show that the data from WC
models may have higher coverage and diversity, but also exhibit higher false pos-
itive rates. We then finetune LMs on data from SE and WC models in different
settings: knowledge distillation, self-improvement, and a novel weak-to-strong
improvement setup where a weaker LM teaches reasoning to a stronger LM. Our
findings reveal that models finetuned on WC-generated data consistently outper-
form those trained on SE-generated data across multiple benchmarks and multiple
choices of WC and SE models. These results challenge the prevailing practice of
relying on SE models for synthetic data generation, suggesting that WC may be
the compute-optimal approach for training advanced LM reasoners.
(a) Finetuning LMs with Gemma2 data.
(b) Finetuning LMs with Gemini 1.5 data.
Figure 1: Summary of the results. (a) We finetune Gemma-7B, Gemma2-9B, and Gemma2-27B on
the synthetic data collected from a stronger but more expensive LM (Gemma2-27B) and a weaker
but cheaper LM (Gemma2-9B) in a compute-matched setup for the MATH dataset. We find that
training with Gemma2-9B data is more compute-optimal across diverse finetuning paradigms –
knowledge distillation, self-improvement, and weak-to-strong improvement (i.e. using a weaker
model to improve a stronger model). (b) We finetune Gemma models (7B/9B/27B) on synthetic
data generated by Gemini-1.5-Pro and Gemini-1.5-Flash in a price-matched setup. We find that
finetuning with Flash-generated data consistently outperforms Pro-generated data.
∗Llama experiments in this paper were conducted only by parties outside of Google. Authors affiliated with
Google were responsible for the Gemini and Gemma experiments.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 0