sweelol/gemma3-270m-pruned-baseline-50pc
Model Description
This model is part of the Sweelol AI Hub collection, resulting from experiments in efficient fine-tuning and knowledge distillation on the Gemma-3-270m architecture using the Databricks Dolly-15k dataset on Kaggle TPUs/GPUs.
Key Details:
- Base Model:
google/gemma-3-270m
- Training Data: Databricks Dolly-15k (subset)
Model Card for sweelol/gemma3-270m-pruned
This model is part of the Sweelol AI Hub collection, resulting from experiments in efficient fine-tuning and knowledge distillation on the Gemma-3-270m architecture using the Databricks Dolly-15k dataset on Kaggle TPUs/GPUs.
Full Research Notebook & Benchmark Results: [Coming soon]
Model Details
Model Description
This is a pruned baseline version of the google/gemma-3-270m
model. It was created by applying magnitude-based pruning to remove 50% of the model's weights. Crucially, this specific variant was not subsequently fine-tuned on a specific task or distilled. It represents the pruned model architecture before any task-specific adaptation.
Developed by: Swee.LOL ai
Shared by: Swee.LOL ai
Model type: Causal Language Model (Pruned)
Language(s) (NLP): English
License: Apache-2.0
Finetuned from model: google/gemma-3-270m
Direct Use
This model can be used for general text generation tasks. It is intended for research and experimentation, particularly in the areas of model compression and the impact of pruning on model performance.
This model is ideally suited as a starting point for further experimentation, such as:
- Fine-tuning on specific instruction-following datasets.
- Using it as a student model in Knowledge Distillation experiments.
- Applying Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA or Prompt Tuning.
Out-of-Scope Use
This model is not intended for production deployment without further fine-tuning. It should not be used for generating content that requires high factual accuracy, safety, or alignment with specific ethical guidelines, as it has not been specifically aligned for these purposes.
Bias, Risks, and Limitations
- Inherent Biases: The model inherits all biases present in the base
google/gemma-3-270m
model and thedatabricks/databricks-dolly-15k
dataset. - Reduced Capability: The aggressive 50% pruning may have reduced the model's general language understanding and reasoning capabilities compared to the full model, as evidenced by lower performance on the HellaSwag benchmark.
- Generalization: The model's performance on tasks outside the scope of its training data may be unpredictable.
Recommendations
Users should be aware that while this model performed surprisingly well on MMLU benchmarks in our experiments, it is a research artifact. It should be evaluated thoroughly on any target task before deployment. The trade-off between model size/efficiency and general capability should be carefully considered.
How to Get Started with the Model
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "sweelol/gemma3-270m-pruned"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)
prompt = "Explain the concept of model pruning."
inputs = tokenizer(prompt, return_tensors="pt")
# Generate
generate_ids = model.generate(inputs.input_ids, max_length=30)
result = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
print(result).
Training Details
Pruning Data
The databricks/databricks-dolly-15k dataset was used to score the importance of weights in the google/gemma-3-270m model to determine which ones to prune. This specific version was not trained on additional data after pruning.
Pruning Procedure
- Sparsity Level: 50%
- Method: Unstructured, magnitude-based pruning.
Evaluation
Testing Data & Metrics
The model was evaluated on a comprehensive suite of tasks from the lm-evaluation-harness
, including 5 diverse subsets of MMLU (for academic reasoning) and HellaSwag (for common-sense reasoning). The primary metric is zero-shot accuracy on a 200-sample subset of each task's test split.
Results
Evaluation
Testing Data & Metrics
The model was evaluated on a comprehensive suite of tasks from the lm-evaluation-harness
, including 5 diverse subsets of MMLU (for academic reasoning) and HellaSwag (for common-sense reasoning). The primary metric is zero-shot accuracy on a 200-sample subset of each task's test split.
Results
This table summarizes the final benchmark scores for the sweelol/gemma3-270m-pruned-baseline
model. It is compared directly against the original, un-pruned google/gemma-3-270m
to measure the impact of pruning.
Benchmark Task | Sweelol Pruned-Baseline | Baseline (Gemma-3-270m) | Change |
---|---|---|---|
Average MMLU (5 tasks) | 26.17% | 24.88% | +1.29% |
HellaSwag (Common Sense) | 29.50% | 43.50% | -14.00% |
---------------------------------- | ---------- | ---------- | -------- |
MMLU Sub-task Breakdown: | |||
MMLU - High School Computer Science | 28.00% | 24.00% | +4.00% |
MMLU - Formal Logic | 29.37% | 25.40% | +3.97% |
MMLU - Professional Law | 26.00% | 27.00% | -1.00% |
MMLU - High School Mathematics | 24.50% | 26.00% | -1.50% |
MMLU - Abstract Algebra | 23.00% | 22.00% | +1.00% |
Summary of Findings: The Unreasonable Effectiveness of Pruning
- A Groundbreaking Result: This is the most significant finding of our study. Simply pruning 50% of the weights from the base
Gemma-3-270m
model, with no subsequent fine-tuning, resulted in a significant improvement in average MMLU performance. - Specialization in Logic: The performance gains were most dramatic in the areas of Formal Logic and Computer Science, suggesting that pruning may remove redundant or noisy pathways, forcing the model to rely on its more robust, core reasoning capabilities.
- The Trade-Off: This improvement in logical reasoning came at a significant cost to the model's common-sense ability, as shown by the large drop in the
HellaSwag
score. This highlights a fascinating trade-off between specialized reasoning and general knowledge.
Full comparative results with other fine-tuning techniques can be found in our main research notebook linked at the top of this card. These results are from the final benchmark run, which can be reproduced in our public research notebook. The surprising strength of this pruned-only model, particularly on logic and computer science tasks, is a key finding of our study.
Model Card Authors
SweeLOL-ai
From Google
Gemma 3 model card
Model Page: Gemma
Resources and Technical Documentation:
- [Gemma 3 Technical Report][g3-tech-report]
- [Responsible Generative AI Toolkit][rai-toolkit]
- [Gemma on Kaggle][kaggle-gemma]
- [Gemma on Vertex Model Garden][vertex-mg-gemma3]
Terms of Use: [Terms][terms]
Authors: Google DeepMind
Inputs and outputs
Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 896 x 896 resolution and encoded to 256 tokens each, for the 4B, 12B, and 27B sizes.
- Total input context of 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B and 270M sizes.
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output context up to 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B and 270M sizes per request, subtracting the request input tokens
Citation
@article{gemma_2025,
title={Gemma 3},
url={https://arxiv.org/abs/2503.19786},
publisher={Google DeepMind},
author={Gemma Team},
year={2025}
}
Evaluation Results
This table compares the performance of this Pruned-Baseline model against the original, un-pruned google/gemma-3-270m
.
Benchmark Task | Sweelol Pruned-Baseline | Baseline (Gemma-3-270m) | Change |
---|---|---|---|
Average MMLU (5 tasks) | 26.17% | 24.88% | +1.29% |
HellaSwag (Common Sense) | 29.50% | 43.50% | -14.00% |
---------------------------------- | ---------- | ---------- | -------- |
MMLU Sub-task Breakdown: | |||
MMLU - Formal Logic | 29.37% | 25.40% | +3.97% |
MMLU - High School Computer Science | 28.00% | 24.00% | +4.00% |
MMLU - Professional Law | 26.00% | 27.00% | -1.00% |
MMLU - High School Mathematics | 24.50% | 26.00% | -1.50% |
MMLU - Abstract Algebra | 23.00% | 22.00% | +1.00% |
Summary of Findings
A key finding of our research: simply pruning 50% of the weights significantly improved the model's average MMLU performance, especially in logical reasoning tasks. This came at the cost of its common-sense reasoning ability (HellaSwag).
Gemma 3 model card
Model Page: Gemma
Resources and Technical Documentation:
- [Gemma 3 Technical Report][g3-tech-report]
- [Responsible Generative AI Toolkit][rai-toolkit]
- [Gemma on Kaggle][kaggle-gemma]
- [Gemma on Vertex Model Garden][vertex-mg-gemma3]
Terms of Use: [Terms][terms]
Authors: Google DeepMind
Model Information
Summary description and brief definition of inputs and outputs.
Description
Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma 3 models are multimodal, handling text and image input and generating text output, with open weights for both pre-trained variants and instruction-tuned variants. Gemma 3 has a large, 128K context window, multilingual support in over 140 languages, and is available in more sizes than previous versions. Gemma 3 models are well-suited for a variety of text generation and image understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.
Inputs and outputs
Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 896 x 896 resolution and encoded to 256 tokens each, for the 4B, 12B, and 27B sizes.
- Total input context of 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B and 270M sizes.
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output context up to 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B and 270M sizes per request, subtracting the request input tokens
Citation
@article{gemma_2025,
title={Gemma 3},
url={https://arxiv.org/abs/2503.19786},
publisher={Google DeepMind},
author={Gemma Team},
year={2025}
}
Model Data
Data used for model training and how the data was processed.
Training Dataset
These models were trained on a dataset of text data that includes a wide variety of sources. The 27B model was trained with 14 trillion tokens, the 12B model was trained with 12 trillion tokens, 4B model was trained with 4 trillion tokens, the 1B with 2 trillion tokens, and the 270M with 6 trillion tokens. The knowledge cutoff date for the training data was August 2024. Here are the key components:
- Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages.
- Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions.
- Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.
- Images: A wide range of images enables the model to perform image analysis and visual data extraction tasks.
The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.
Data Preprocessing
Here are the key data cleaning and filtering methods applied to the training data:
- CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content.
- Sensitive Data Filtering: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets.
- Additional methods: Filtering based on content quality and safety in line with [our policies][safety-policies].
Implementation Information
Details about the model internals.
Hardware
Gemma was trained using [Tensor Processing Unit (TPU)][tpu] hardware (TPUv4p, TPUv5p and TPUv5e). Training vision-language models (VLMS) requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:
- Performance: TPUs are specifically designed to handle the massive computations involved in training VLMs. They can speed up training considerably compared to CPUs.
- Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training. This can lead to better model quality.
- Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
- Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU-based infrastructure, especially when considering the time and resources saved due to faster training.
- These advantages are aligned with [Google's commitments to operate sustainably][sustainability].
Software
Training was done using [JAX][jax] and [ML Pathways][ml-pathways].
JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is Google's latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like these ones.
Together, JAX and ML Pathways are used as described in the [paper about the Gemini family of models][gemini-2-paper]; "the 'single controller' programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow."
Evaluation
Model evaluation metrics and results.
Benchmark Results
These models were evaluated against a large collection of different datasets and metrics to cover different aspects of text generation. Evaluation results marked with IT are for instruction-tuned models. Evaluation results marked with PT are for pre-trained models.
Gemma 3 270M
Benchmark | n-shot | Gemma 3 PT 270M |
---|---|---|
HellaSwag | 10-shot | 40.9 |
BoolQ | 0-shot | 61.4 |
PIQA | 0-shot | 67.7 |
TriviaQA | 5-shot | 15.4 |
ARC-c | 25-shot | 29.0 |
ARC-e | 0-shot | 57.7 |
WinoGrande | 5-shot | 52.0 |
Benchmark | n-shot | Gemma 3 IT 270m |
---|---|---|
HellaSwag | 0-shot | 37.7 |
PIQA | 0-shot | 66.2 |
ARC-c | 0-shot | 28.2 |
WinoGrande | 0-shot | 52.3 |
BIG-Bench Hard | few-shot | 26.7 |
IF Eval | 0-shot | 51.2 |
Gemma 3 1B, 4B, 12B & 27B
Reasoning and factuality
Benchmark | n-shot | Gemma 3 IT 1B | Gemma 3 IT 4B | Gemma 3 IT 12B | Gemma 3 IT 27B |
---|---|---|---|---|---|
GPQA Diamond | 0-shot | 19.2 | 30.8 | 40.9 | 42.4 |
SimpleQA | 0-shot | 2.2 | 4.0 | 6.3 | 10.0 |
FACTS Grounding | - | 36.4 | 70.1 | 75.8 | 74.9 |
BIG-Bench Hard | 0-shot | 39.1 | 72.2 | 85.7 | 87.6 |
BIG-Bench Extra Hard | 0-shot | 7.2 | 11.0 | 16.3 | 19.3 |
IFEval | 0-shot | 80.2 | 90.2 | 88.9 | 90.4 |
Benchmark | n-shot | Gemma 3 PT 1B | Gemma 3 PT 4B | Gemma 3 PT 12B | Gemma 3 PT 27B |
---|---|---|---|---|---|
HellaSwag | 10-shot | 62.3 | 77.2 | 84.2 | 85.6 |
BoolQ | 0-shot | 63.2 | 72.3 | 78.8 | 82.4 |
PIQA | 0-shot | 73.8 | 79.6 | 81.8 | 83.3 |
SocialIQA | 0-shot | 48.9 | 51.9 | 53.4 | 54.9 |
TriviaQA | 5-shot | 39.8 | 65.8 | 78.2 | 85.5 |
Natural Questions | 5-shot | 9.48 | 20.0 | 31.4 | 36.1 |
ARC-c | 25-shot | 38.4 | 56.2 | 68.9 | 70.6 |
ARC-e | 0-shot | 73.0 | 82.4 | 88.3 | 89.0 |
WinoGrande | 5-shot | 58.2 | 64.7 | 74.3 | 78.8 |
BIG-Bench Hard | few-shot | 28.4 | 50.9 | 72.6 | 77.7 |
DROP | 1-shot | 42.4 | 60.1 | 72.2 | 77.2 |
STEM and code
Benchmark | n-shot | Gemma 3 IT 1B | Gemma 3 IT 4B | Gemma 3 IT 12B | Gemma 3 IT 27B |
---|---|---|---|---|---|
[MMLU][mmlu] (Pro) | 0-shot | 14.7 | 43.6 | 60.6 | 67.5 |
[LiveCodeBench][lcb] | 0-shot | 1.9 | 12.6 | 24.6 | 29.7 |
[Bird-SQL][bird-sql] (dev) | - | 6.4 | 36.3 | 47.9 | 54.4 |
[Math][math] | 0-shot | 48.0 | 75.6 | 83.8 | 89.0 |
HiddenMath | 0-shot | 15.8 | 43.0 | 54.5 | 60.3 |
[MBPP][mbpp] | 3-shot | 35.2 | 63.2 | 73.0 | 74.4 |
[HumanEval][humaneval] | 0-shot | 41.5 | 71.3 | 85.4 | 87.8 |
[Natural2Code][nat2code] | 0-shot | 56.0 | 70.3 | 80.7 | 84.5 |
[GSM8K][gsm8k] | 0-shot | 62.8 | 89.2 | 94.4 | 95.9 |
Benchmark | n-shot | Gemma 3 PT 4B | Gemma 3 PT 12B | Gemma 3 PT 27B |
---|---|---|---|---|
[MMLU][mmlu] | 5-shot | 59.6 | 74.5 | 78.6 |
[MMLU][mmlu] (Pro COT) | 5-shot | 29.2 | 45.3 | 52.2 |
[AGIEval][agieval] | 3-5-shot | 42.1 | 57.4 | 66.2 |
[MATH][math] | 4-shot | 24.2 | 43.3 | 50.0 |
[GSM8K][gsm8k] | 8-shot | 38.4 | 71.0 | 82.6 |
GPQA | 5-shot | 15.0 | 25.4 | 24.3 |
[MBPP][mbpp] | 3-shot | 46.0 | 60.4 | 65.6 |
[HumanEval][humaneval] | 0-shot | 36.0 | 45.7 | 48.8 |
Model tree for sweelol/gemma3-270m-pruned-baseline-50pc
Base model
google/gemma-3-270m