TokSuite: Measuring the Impact of Tokenizer Choice on Language Model Behavior
Paper
•
2512.20757
•
Published
•
14
AI & ML interests
Exploring smol models (for text, vision and video) and high quality web and synthetic datasets
Recent Activity
View all activity
Papers
View all Papers
craffel
authored
a
paper
2 days ago
qgallouedec
submitted
a
paper
to Daily Papers
4 days ago
loubnabnl
updated
a
dataset
6 days ago
The number of total tokens in the comment seems incorrect
👍
1
1
#2 opened 8 days ago
by
tianjun
loubnabnl
updated
a
Space
6 days ago
eliebak
submitted
a
paper
to Daily Papers
9 days ago
Inconsistent numbers
6
#1 opened about 1 month ago
by
MysticJay
alozowski
authored
a
paper
19 days ago
Post
1310
🚀🚀🚀Huge biotech data drop today🚀🚀🚀
The largest drug-target dataset ever created was just released on Hugging Face—and it's still growing...
EvE Bio is further updating the dataset every 8 weeks. Drug development dream.
Read the blog: https://huggingface.co/blog/hugging-science/eve-bio-mapping-the-pharmone-drug-interaction
Play with the data: eve-bio/drug-target-activity
The largest drug-target dataset ever created was just released on Hugging Face—and it's still growing...
EvE Bio is further updating the dataset every 8 weeks. Drug development dream.
Read the blog: https://huggingface.co/blog/hugging-science/eve-bio-mapping-the-pharmone-drug-interaction
Play with the data: eve-bio/drug-target-activity
nouamanetazi
posted
an
update
about 2 months ago
Post
4124
After training 𝐒𝐦𝐨𝐥𝐋𝐌𝟑 on 𝟑𝟖𝟒 𝐇𝟏𝟎𝟎𝐬 for nearly a month, I've come to realize something most people overlook: 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐦𝐚𝐤𝐞-𝐨𝐫-𝐛𝐫𝐞𝐚𝐤 𝐟𝐚𝐜𝐭𝐨𝐫 𝐢𝐧 𝐋𝐋𝐌 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠. 🔥
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious 𝐍𝐂𝐂𝐋 𝐞𝐫𝐫𝐨𝐫𝐬, or when your expensive GPU cluster is running at 𝟔𝟎% 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲, the problem isn't your model. It's most probably a 𝐦𝐢𝐬𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 🛠️
Questions that seemed simple but had no clear answers: Why is 𝐌𝐨𝐄 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐬𝐥𝐨𝐰𝐞𝐫 𝐭𝐡𝐚𝐧 𝐝𝐞𝐧𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬? Which 𝐍𝐂𝐂𝐋 𝐟𝐥𝐚𝐠𝐬 should we actually set? How often should we checkpoint without killing throughput?
That's why we built 𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤 📖: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐥𝐚𝐲𝐞𝐫 that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: 𝐇𝐁𝐌𝟑 𝐡𝐢𝐭𝐭𝐢𝐧𝐠 𝟑 𝐓𝐁/𝐬, 𝐍𝐕𝐋𝐢𝐧𝐤 𝟒.𝟎 𝐫𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝟕𝟖𝟔 𝐆𝐁/𝐬, 𝐏𝐂𝐈𝐞 𝐆𝐞𝐧𝟒 𝐚𝐭 𝟏𝟒.𝟐 𝐆𝐁/𝐬. Then we ran collective operations across 𝟏𝟐𝟖 𝐆𝐏𝐔𝐬 (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from 𝟒𝟖𝟎 𝐆𝐁/𝐬 on a single node to 𝟑𝟐𝟎-𝟑𝟓𝟎 𝐆𝐁/𝐬 across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: https://lnkd.in/e5MKXUHS
Shared with ❤️ by the HuggingFace team
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious 𝐍𝐂𝐂𝐋 𝐞𝐫𝐫𝐨𝐫𝐬, or when your expensive GPU cluster is running at 𝟔𝟎% 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲, the problem isn't your model. It's most probably a 𝐦𝐢𝐬𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 🛠️
Questions that seemed simple but had no clear answers: Why is 𝐌𝐨𝐄 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐬𝐥𝐨𝐰𝐞𝐫 𝐭𝐡𝐚𝐧 𝐝𝐞𝐧𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬? Which 𝐍𝐂𝐂𝐋 𝐟𝐥𝐚𝐠𝐬 should we actually set? How often should we checkpoint without killing throughput?
That's why we built 𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤 📖: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐥𝐚𝐲𝐞𝐫 that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: 𝐇𝐁𝐌𝟑 𝐡𝐢𝐭𝐭𝐢𝐧𝐠 𝟑 𝐓𝐁/𝐬, 𝐍𝐕𝐋𝐢𝐧𝐤 𝟒.𝟎 𝐫𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝟕𝟖𝟔 𝐆𝐁/𝐬, 𝐏𝐂𝐈𝐞 𝐆𝐞𝐧𝟒 𝐚𝐭 𝟏𝟒.𝟐 𝐆𝐁/𝐬. Then we ran collective operations across 𝟏𝟐𝟖 𝐆𝐏𝐔𝐬 (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from 𝟒𝟖𝟎 𝐆𝐁/𝐬 on a single node to 𝟑𝟐𝟎-𝟑𝟓𝟎 𝐆𝐁/𝐬 across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: https://lnkd.in/e5MKXUHS
Shared with ❤️ by the HuggingFace team
andito
authored
a
paper
2 months ago
Post
1883
Finally, our new paper is out! "𝗙𝗶𝗻𝗲𝗩𝗶𝘀𝗶𝗼𝗻: 𝗢𝗽𝗲𝗻 𝗗𝗮𝘁𝗮 𝗜𝘀 𝗔𝗹𝗹 𝗬𝗼𝘂 𝗡𝗲𝗲𝗱"! 🥳
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
🔍 finding and cleaning data at scale
🧹 removing excessive duplicates across sources
🤗 decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
🎉 To celebrate the paper, I’m also releasing a concatenated and shuffled version of the full dataset! 👉
It’s ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
🔍 finding and cleaning data at scale
🧹 removing excessive duplicates across sources
🤗 decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
🎉 To celebrate the paper, I’m also releasing a concatenated and shuffled version of the full dataset! 👉
HuggingFaceM4/FineVision_full_shuffled It’s ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
Post
7348
deepseek-ai/DeepSeek-OCR is out! 🔥 my take ⤵️
> pretty insane it can parse and re-render charts in HTML
> it uses CLIP and SAM features concatenated, so better grounding
> very efficient per vision tokens/performance ratio
> covers 100 languages
> pretty insane it can parse and re-render charts in HTML
> it uses CLIP and SAM features concatenated, so better grounding
> very efficient per vision tokens/performance ratio
> covers 100 languages
- 4 replies
thomwolf
authored
a
paper
2 months ago
lvwerra
authored
a
paper
3 months ago
sasha
authored
3
papers
3 months ago
Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model
Paper
•
2211.02001
•
Published
Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI
Paper
•
2409.14160
•
Published
•
3
From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI's Polarized Environmental Debate
Paper
•
2501.16548
•
Published
Post
3285
🚀 New blog: Maintain the unmaintainable – 1M+ Python LOC, 400+ models
How do you stop a million-line library built by thousands of contributors from collapsing under its own weight?
At 🤗 Transformers, we do it with explicit software-engineering tenets, principles that make the codebase hackable at scale.
🔍 Inside the post:
– One Model, One File: readability first — you can still open a modeling file and see the full logic, top to bottom.
– Modular Transformers: visible inheritance that cuts maintenance cost by ~15× while keeping models readable.
– Config-Driven Performance: FlashAttention, tensor parallelism, and attention scheduling are config-level features, not rewrites.
Written with @lysandre ,@pcuenq and @yonigozlan , this is a deep dive into how Transformers stays fast, open, and maintainable.
Read it here → transformers-community/Transformers-tenets
How do you stop a million-line library built by thousands of contributors from collapsing under its own weight?
At 🤗 Transformers, we do it with explicit software-engineering tenets, principles that make the codebase hackable at scale.
🔍 Inside the post:
– One Model, One File: readability first — you can still open a modeling file and see the full logic, top to bottom.
– Modular Transformers: visible inheritance that cuts maintenance cost by ~15× while keeping models readable.
– Config-Driven Performance: FlashAttention, tensor parallelism, and attention scheduling are config-level features, not rewrites.
Written with @lysandre ,@pcuenq and @yonigozlan , this is a deep dive into how Transformers stays fast, open, and maintainable.
Read it here → transformers-community/Transformers-tenets
Post
6802
large AI labs open-sourced a ton of models last week 🔥
here's few picks, find even more here merve/sep-16-releases-68d13ea4c547f02f95842f05 🤝
> IBM released a new Docling model with 258M params based on Granite (A2.0) 📝 ibm-granite/granite-docling-258M
> Xiaomi released 7B audio LM with base and instruct variants (MIT) XiaomiMiMo/mimo-audio-68cc7202692c27dae881cce0
> DecartAI released Lucy Edit, open Nano Banana 🍌 (NC) decart-ai/Lucy-Edit-Dev
> OpenGVLab released a family of agentic computer use models (3B/7B/32B) with the dataset 💻 OpenGVLab/scalecua-68c912cf56f7ff4c8e034003
> Meituan Longcat released thinking version of LongCat-Flash 💭 meituan-longcat/LongCat-Flash-Thinking
here's few picks, find even more here merve/sep-16-releases-68d13ea4c547f02f95842f05 🤝
> IBM released a new Docling model with 258M params based on Granite (A2.0) 📝 ibm-granite/granite-docling-258M
> Xiaomi released 7B audio LM with base and instruct variants (MIT) XiaomiMiMo/mimo-audio-68cc7202692c27dae881cce0
> DecartAI released Lucy Edit, open Nano Banana 🍌 (NC) decart-ai/Lucy-Edit-Dev
> OpenGVLab released a family of agentic computer use models (3B/7B/32B) with the dataset 💻 OpenGVLab/scalecua-68c912cf56f7ff4c8e034003
> Meituan Longcat released thinking version of LongCat-Flash 💭 meituan-longcat/LongCat-Flash-Thinking
- 2 replies