TokSuite: Measuring the Impact of Tokenizer Choice on Language Model Behavior
Paper
•
2512.20757
•
Published
•
16
FineData
community
AI & ML interests
We release large pre-training datasets to accelerate open LLM development. Part of the Hugging Face Science team (hf.co/science)
Recent Activity
View all activity
Papers
FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language
The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
craffel
authored
a
paper
5 days ago
eliebak
submitted
a
paper
to Daily Papers
12 days ago
Which language detector did you use
4
#28 opened 16 days ago
by
ming030890
The "file_path" data field appears to primarily contain cc-index paths rather than WARC paths.
9
#16 opened 3 months ago
by
lnstrument
A Few Questions About the Implementation Details of the finepdfs Project
5
#24 opened 3 months ago
by
yoliax
Dataset broken by latest update?
5
#27 opened about 1 month ago
by
Rijgersberg
hynky
updated
a
dataset
28 days ago
Post
3848
🤖 Did you know your voice might be cloned without your consent from just *one sentence* of audio?
That's not great. So with @frimelle , we brainstormed a new idea for developers who want to curb malicious use: ✨The Voice Consent Gate.✨
Details, code, here: https://huggingface.co/blog/voice-consent-gate
That's not great. So with @frimelle , we brainstormed a new idea for developers who want to curb malicious use: ✨The Voice Consent Gate.✨
Details, code, here: https://huggingface.co/blog/voice-consent-gate
- 3 replies
thomwolf
authored
a
paper
3 months ago
lvwerra
authored
a
paper
3 months ago
Post
2909
🤖 As AI-generated content is shared in movies/TV/across the web, there's one simple low-hanging fruit 🍇 to help know what's real: Visible watermarks. With the Gradio team, I've made sure it's trivially easy to add this disclosure to images, video, chatbot text. See how: https://huggingface.co/blog/watermarking-with-gradio
Thanks to the code collab in particular from @abidlabs and Yuvraj Sharma.
Thanks to the code collab in particular from @abidlabs and Yuvraj Sharma.
davanstrien
posted
an
update
4 months ago
Post
1492
I fine-tuned a smol VLM to generate specialized art history metadata!
https://huggingface.co/davanstrien/iconclass-vlm: Qwen2.5-VL-3B trained using SFT to generate ICONCLASS codes (think Dewey Decimal for art!)
Trained with TRL + HF Jobs - single UV script, no GPU needed!
Space to explore predictions on a test set: davanstrien/iconclass-predictions
Blog soon!
https://huggingface.co/davanstrien/iconclass-vlm: Qwen2.5-VL-3B trained using SFT to generate ICONCLASS codes (think Dewey Decimal for art!)
Trained with TRL + HF Jobs - single UV script, no GPU needed!
Space to explore predictions on a test set: davanstrien/iconclass-predictions
Blog soon!
Post
3838
Super excited to announce that our research team at Hugging Face will be doing an AMA on reddit r/LocalLLaMA.
Come ask any questions to the team behind SmolLM, FineWeb and more! And who knows, maybe there’ll be a shiny new release to talk about?
Thursday 4th September, 8AM-11AM PST 🤗
science
Come ask any questions to the team behind SmolLM, FineWeb and more! And who knows, maybe there’ll be a shiny new release to talk about?
Thursday 4th September, 8AM-11AM PST 🤗
Post
709
Motif 2.6B tech report is pretty insane, first time i see a model with differential attention and polynorm trained at scale!
> It's trained on 2.5T of token, with a "data mixture schedule" to continuously adjust the mixture over training.
> They use WSD with a "Simple moving average" averaging the last 6 ckpt every 8B token.
> They trained on Finemath, Fineweb2, DCLM, TxT360.
> Lot of details in the finetuning data they used, for instance they used EvolKit and did some "dataset fusion" to have more compressed knowledge into the data.
> They mention they also tried Normalized GPT, QK-Norm and Cross Layer Attention.
Motif-Technologies/Motif-2.6B
> It's trained on 2.5T of token, with a "data mixture schedule" to continuously adjust the mixture over training.
> They use WSD with a "Simple moving average" averaging the last 6 ckpt every 8B token.
> They trained on Finemath, Fineweb2, DCLM, TxT360.
> Lot of details in the finetuning data they used, for instance they used EvolKit and did some "dataset fusion" to have more compressed knowledge into the data.
> They mention they also tried Normalized GPT, QK-Norm and Cross Layer Attention.
Motif-Technologies/Motif-2.6B
Post
2965
New work from my socially-minded colleagues at Hugging Face, creating some foundations for AI companionship behavior evaluation.
Evaluation Dataset: AI-companionship/INTIMA
Paper: AI-companionship/INTIMA
Work from @giadap , @frimelle , @yjernite .
Evaluation Dataset: AI-companionship/INTIMA
Paper: AI-companionship/INTIMA
Work from @giadap , @frimelle , @yjernite .
- 2 replies
Post
453
🤖 ICYMI: Yesterday, Hugging Face and OpenAI partnered to bring open source GPT to the public. This is a Big Deal in "AI world".
0. Common ground setting: OpenAI is the ChatGPT people. An “open source” model is one whose weights are available — that means the model can be “yours”.
1. You don’t have to interact with the company directly, nor give them your interactions, to use the system. The company can't "surveil" you.
2. You can evaluate the unique contributions of their SOTA model much more rigorously than you can when there are collections of models+code behind a closed API. You can find out specifically what the model can and can't do.
3. And you can directly customize it for whatever you'd like. Fine-tuning, wherein you give the model data that's tailored to your use cases and train it some more on that data, is trivial* when you have the model weights.
*Provided you have the compute.
4. You can directly benchmark whatever you'd like. Biases? Energy usage? Strengths/weaknesses? Go for it. You wants it you gots it--this transparency helps people understand SOTA *in general*, not just for this model, but points to, e.g., what's going on with closed Google models as well.
5. One of the most powerful things about "openness" that I've learned is that it cultivates ecosystems of collaborators building on top of one another's brilliance to make systems that are significantly better than they would be if created in isolation.
But, caveat wrt my own philosophy...
6. I do not take it as a given that advancing LLMs is good, and have a lot more to say wrt where I think innovation should focus more. For example, a focus on *data* -- curation, measurement, consent, credit, compensation, safety -- would deeply improve technology for everyone.
7. The transparency this release provides is massive for people who want to *learn* about LLMs. For the next generation of technologists to advance over the current, they MUST be able to learn about what's happening now. (cont...)
0. Common ground setting: OpenAI is the ChatGPT people. An “open source” model is one whose weights are available — that means the model can be “yours”.
1. You don’t have to interact with the company directly, nor give them your interactions, to use the system. The company can't "surveil" you.
2. You can evaluate the unique contributions of their SOTA model much more rigorously than you can when there are collections of models+code behind a closed API. You can find out specifically what the model can and can't do.
3. And you can directly customize it for whatever you'd like. Fine-tuning, wherein you give the model data that's tailored to your use cases and train it some more on that data, is trivial* when you have the model weights.
*Provided you have the compute.
4. You can directly benchmark whatever you'd like. Biases? Energy usage? Strengths/weaknesses? Go for it. You wants it you gots it--this transparency helps people understand SOTA *in general*, not just for this model, but points to, e.g., what's going on with closed Google models as well.
5. One of the most powerful things about "openness" that I've learned is that it cultivates ecosystems of collaborators building on top of one another's brilliance to make systems that are significantly better than they would be if created in isolation.
But, caveat wrt my own philosophy...
6. I do not take it as a given that advancing LLMs is good, and have a lot more to say wrt where I think innovation should focus more. For example, a focus on *data* -- curation, measurement, consent, credit, compensation, safety -- would deeply improve technology for everyone.
7. The transparency this release provides is massive for people who want to *learn* about LLMs. For the next generation of technologists to advance over the current, they MUST be able to learn about what's happening now. (cont...)
- 1 reply
Post
505
🤖 👾 Thanks so much to BBC News and the stellar Suranjana Tewari for having me on to talk about US <—> China relationship in AI, and what it means for AI ethics.
Post
4788
Kimi K2 tech report is full of gems as always. Here are my notes on it:
> MuonClip: Pretty crazy how after 70k the training stabilizes and the QK-clip is basically inactive. There is also no loss in perf with QK-clip which is not trivial at all (at small scale but with aggressive threshold). Also a cool explanation of why muon makes the logit explode in appendix E (tl;dr is that muon makes the singular value of the update matrix higher)
> Sparsity scaling laws to justify their ratio, they have a very solid training infra that allows the model to be trained at this sparsity level, they could have increased even more but as sparsity increases the training becomes less efficient.
> They diminish the number of attention heads to make it more efficient for long context since attention heads are a big bottleneck for long context. They also remove 2 of the 3 "first dense" layers in the dsv3 arch.
With the sparsity and attention heads (divided by 2) they achieve 83% increased flops compared to deepseek v3 arch at 128k.
> Data: Rephrasing is KEY. They do a lot more synthetic data generation and rephrase their corpus to have different styles, for longer documents they do it by chunk. I'm (half) surprised by the fact that ONLY 1 epoch (assuming same number of training tokens I think?) of data rephrased 10 times has better accuracy than 10 epochs of the same data rephrased once.
> They do rewriting for Math and Knowledge, for Math they apply the ShallowMath recipe and instruct the model to rephrase in a "learning note" style
> They talk about diversity and probably have some internal stuff/eval to test that, as always still a bit unclear for me how to properly measure that.
The infra is also very nice, quick summary:
> PP=16 (1F1B schedule, a bit custom), EP=16, zero1
> No FP8 computation but for storage of specific layers, selective recomputation for inexpensive block, activation offloading to CPU
> MuonClip: Pretty crazy how after 70k the training stabilizes and the QK-clip is basically inactive. There is also no loss in perf with QK-clip which is not trivial at all (at small scale but with aggressive threshold). Also a cool explanation of why muon makes the logit explode in appendix E (tl;dr is that muon makes the singular value of the update matrix higher)
> Sparsity scaling laws to justify their ratio, they have a very solid training infra that allows the model to be trained at this sparsity level, they could have increased even more but as sparsity increases the training becomes less efficient.
> They diminish the number of attention heads to make it more efficient for long context since attention heads are a big bottleneck for long context. They also remove 2 of the 3 "first dense" layers in the dsv3 arch.
With the sparsity and attention heads (divided by 2) they achieve 83% increased flops compared to deepseek v3 arch at 128k.
> Data: Rephrasing is KEY. They do a lot more synthetic data generation and rephrase their corpus to have different styles, for longer documents they do it by chunk. I'm (half) surprised by the fact that ONLY 1 epoch (assuming same number of training tokens I think?) of data rephrased 10 times has better accuracy than 10 epochs of the same data rephrased once.
> They do rewriting for Math and Knowledge, for Math they apply the ShallowMath recipe and instruct the model to rephrase in a "learning note" style
> They talk about diversity and probably have some internal stuff/eval to test that, as always still a bit unclear for me how to properly measure that.
The infra is also very nice, quick summary:
> PP=16 (1F1B schedule, a bit custom), EP=16, zero1
> No FP8 computation but for storage of specific layers, selective recomputation for inexpensive block, activation offloading to CPU
NXz64Fdf8Y
authored
a
paper
6 months ago
thomwolf
authored
a
paper
6 months ago