SmolVLM: Redefining small and efficient multimodal models
Paper
β’
2504.05299
β’
Published
β’
158
smol-ablations
Enterprise
AI & ML interests
None defined yet.
Recent Activity
View all activity
smol-ablations's activity
nouamanetaziΒ
authored
a
paper
7 days ago
anton-lΒ
authored
a
paper
7 days ago
eliebakΒ
authored
a
paper
7 days ago
loubnabnlΒ
authored
a
paper
8 days ago
Post
1653
Google just dropped an exciting technical report for the brand-new Gemma3 model! π Here are my personal notes highlighting the most intriguing architectural innovations, design choices, and insights from this release:
1) Architecture choices:
> No more softcaping, replace by QK-Norm
> Both Pre AND Post Norm
> Wider MLP than Qwen2.5, ~ same depth
> SWA with 5:1 and 1024 (very small and cool ablation on the paper!)
> No MLA to save KV cache, SWA do the job!
2) Long context
> Only increase the rope in the global layer (to 1M)
> Confirmation that it's harder to do long context for smol models, no 128k for the 1B
> Pretrained with 32k context? seems very high
> No yarn nor llama3 like rope extension
3) Distillation
> Only keep te first 256 logits for the teacher
> Ablation on the teacher gap (tl;dr you need some "patience" to see that using a small teacher is better)
> On policy distillation yeahh (by
@agarwl_
et al), not sure if the teacher gap behave the same here, curious if someone have more info?
4) Others
> Checkpoint with QAT, that's very cool
> RL using improve version of BOND, WARM/WARP good excuse to look at
@ramealexandre
papers
> Only use Zero3, no TP/PP if i understand correctly ?
> Training budget relatively similar than gemma2
1) Architecture choices:
> No more softcaping, replace by QK-Norm
> Both Pre AND Post Norm
> Wider MLP than Qwen2.5, ~ same depth
> SWA with 5:1 and 1024 (very small and cool ablation on the paper!)
> No MLA to save KV cache, SWA do the job!
2) Long context
> Only increase the rope in the global layer (to 1M)
> Confirmation that it's harder to do long context for smol models, no 128k for the 1B
> Pretrained with 32k context? seems very high
> No yarn nor llama3 like rope extension
3) Distillation
> Only keep te first 256 logits for the teacher
> Ablation on the teacher gap (tl;dr you need some "patience" to see that using a small teacher is better)
> On policy distillation yeahh (by
@agarwl_
et al), not sure if the teacher gap behave the same here, curious if someone have more info?
4) Others
> Checkpoint with QAT, that's very cool
> RL using improve version of BOND, WARM/WARP good excuse to look at
@ramealexandre
papers
> Only use Zero3, no TP/PP if i understand correctly ?
> Training budget relatively similar than gemma2
- 1 reply
eliebakΒ
authored
a
paper
2 months ago
loubnabnlΒ
authored
a
paper
2 months ago
eliebakΒ
authored
a
paper
2 months ago
anton-lΒ
authored
a
paper
2 months ago
Post
2745
Introducing ππ
π’π§πππππ‘: the best public math pre-training dataset with 50B+ tokens!
HuggingFaceTB/finemath
Math remains challenging for LLMs and by training on FineMath we see considerable gains over other math datasets, especially on GSM8K and MATH.
We build the dataset by:
π οΈ carefully extracting math data from Common Crawl;
π iteratively filtering and recalling high quality math pages using a classifier trained on synthetic annotations to identify math reasoning and deduction.
We conducted a series of ablations comparing the performance of Llama-3.2-3B-Base after continued pre-training on FineMath and observe notable gains compared to the baseline model and other public math datasets.
We hope this helps advance the performance of LLMs on math and reasoning! π
Weβre also releasing all the ablation models as well as the evaluation code.
HuggingFaceTB/finemath-6763fb8f71b6439b653482c2
HuggingFaceTB/finemath
Math remains challenging for LLMs and by training on FineMath we see considerable gains over other math datasets, especially on GSM8K and MATH.
We build the dataset by:
π οΈ carefully extracting math data from Common Crawl;
π iteratively filtering and recalling high quality math pages using a classifier trained on synthetic annotations to identify math reasoning and deduction.
We conducted a series of ablations comparing the performance of Llama-3.2-3B-Base after continued pre-training on FineMath and observe notable gains compared to the baseline model and other public math datasets.
We hope this helps advance the performance of LLMs on math and reasoning! π
Weβre also releasing all the ablation models as well as the evaluation code.
HuggingFaceTB/finemath-6763fb8f71b6439b653482c2
Post
3431
Making SmolLM2 reproducible: open-sourcing our training & evaluation toolkit π οΈ https://github.com/huggingface/smollm/
- Pre-training code with nanotron
- Evaluation suite with lighteval
- Synthetic data generation using distilabel (powers our new SFT dataset HuggingFaceTB/smoltalk)
- Post-training scripts with TRL & the alignment handbook
- On-device tools with llama.cpp for summarization, rewriting & agents
Apache 2.0 licensed. V2 pre-training data mix coming soon!
Which other tools should we add next?
- Pre-training code with nanotron
- Evaluation suite with lighteval
- Synthetic data generation using distilabel (powers our new SFT dataset HuggingFaceTB/smoltalk)
- Post-training scripts with TRL & the alignment handbook
- On-device tools with llama.cpp for summarization, rewriting & agents
Apache 2.0 licensed. V2 pre-training data mix coming soon!
Which other tools should we add next?
anton-lΒ
authored
a
paper
10 months ago
loubnabnlΒ
authored
a
paper
10 months ago
Post
1831
Wow, impressive 340B model by nvidia with a nice permissive license! π The technical report is full of insights and seems to use a different learning rate schedule than cosine, probably a variant of WSD. Hope to get more info on that! π
nvidia/nemotron-4-340b-666b7ebaf1b3867caf2f1911
nvidia/nemotron-4-340b-666b7ebaf1b3867caf2f1911
Post
5885
π· FineWeb technical report is out and so is π FineWeb-Edu, a 1.3 trillion tokens dataset that outperforms all other open web datasets, with remarkable improvements on educational benchmarksΒ such as MMLU, ARC, and OpenBookQA.
Technical report: HuggingFaceFW/blogpost-fineweb-v1
Dataset: HuggingFaceFW/fineweb-edu
We used Llama 3 generations to train an educational quality classifier, filtering the 15 trillion tokens of FineWeb to select only those with high educational value (an approach also used in Llama 3 and Phi-3 training datasets). We're releasing both FineWeb-Edu and the classifier, along with a larger, less heavily filtered version containing 5.4 trillion tokens.
You can find more details about the dataset and the experiments we ran in the FineWeb technical report, It's a 45-minute read but it contains all the secret sauce for building high quality web datasets.
Enjoy!
Technical report: HuggingFaceFW/blogpost-fineweb-v1
Dataset: HuggingFaceFW/fineweb-edu
We used Llama 3 generations to train an educational quality classifier, filtering the 15 trillion tokens of FineWeb to select only those with high educational value (an approach also used in Llama 3 and Phi-3 training datasets). We're releasing both FineWeb-Edu and the classifier, along with a larger, less heavily filtered version containing 5.4 trillion tokens.
You can find more details about the dataset and the experiments we ran in the FineWeb technical report, It's a 45-minute read but it contains all the secret sauce for building high quality web datasets.
Enjoy!
eliebakΒ
authored
a
paper
11 months ago
loubnabnlΒ
authored
a
paper
11 months ago
stillermanΒ
authored
a
paper
about 1 year ago
Post
6438
We've just published a detailed blog post on the creation of Cosmopedia dataset. We hope this will provide insights about generating synthetic data at scale for pre-training.
https://huggingface.co/blog/cosmopedia
Here are some key takeaways:
π― Prompt curation is crucial: we want to cover many topics with few duplicates.
π You can leverage various resources for diversity: using different seed data, generation formats, and target audiences.
βοΈ The importance of a good technical stack: for scalable generations with tools like llm-swarm and fast model training and evaluation.
Have a good read!
https://huggingface.co/blog/cosmopedia
Here are some key takeaways:
π― Prompt curation is crucial: we want to cover many topics with few duplicates.
π You can leverage various resources for diversity: using different seed data, generation formats, and target audiences.
βοΈ The importance of a good technical stack: for scalable generations with tools like llm-swarm and fast model training and evaluation.
Have a good read!
- 1 reply
nouamanetaziΒ
authored
a
paper
about 1 year ago