davanstrien (Daniel van Strien)

posted an update 3 days ago

Post

2327

Hacked together a way to log trl GRPO training completions to a 🤗 dataset repo. This allows you to:

- Track rewards from multiple reward functions
- Treat the completion and rewards from training as a "proper" dataset and do EDA
- Share results for open science

The implementation is super hacky, but I'm curious if people would find this useful.

To push completions to the Hub, you just need two extra parameters:

log_completions=True
log_completions_hub_repo='your-username/repo-name'

Example dataset: davanstrien/test-logs
Colab: https://colab.research.google.com/drive/1wzBFPVthRYYTp-mEYlznLg_e_0Za1M3g

posted an update 7 days ago

Post

2162

Dataset descriptions for trending Hugging Face datasets? Powered by a Smol model davanstrien/Smol-Hub-tldr

posted an update 9 days ago

Post

1836

How do you make 1M+ Hugging Face models & datasets more discoverable?

davanstrien/Smol-Hub-tldr!

I fine-tuned HuggingFaceTB/SmolLM2-360M to generate one-line summaries from a model or dataset README.

Its own self-description?
"A model for generating concise summaries of model & dataset cards from the Hugging Face Hub"

The goal? Make it easier to find the right models and datasets for your specific needs. It's already powering a semantic search for datasets Space.

It's still a WIP but thanks to @loubnabnl , @anton-l , @eliebak et al, for cooking such a nice base model for fine-tuning small, efficient models for specific domains and tasks. 🙏

posted an update 10 days ago

Post

1312

Made some significant updates to my 🤗 semantic datasets search app. If you love falling into a wiki black hole, you might like this...

librarian-bots/huggingface-datasets-semantic-search

reacted to Ihor's post with 🚀 23 days ago

Post

1439

🚀 Reproducing DeepSeek R1 for Text-to-Graph Extraction

I’ve been working on replicating DeepSeek R1, focusing on zero-shot text-to-graph extraction—a challenging task where LMs extract entities and relations from text based on predefined types.

🧠 Key Insight:
Language models struggle when constrained by entity/relation types. Supervised training alone isn’t enough, but reinforcement learning (RL), specifically Guided Reward Policy Optimization (GRPO), shows promise.

💡 Why GRPO?
It trains the model to generate structured graphs, optimizing multiple reward functions (format, JSON validity, and extraction accuracy).
It allows the model to learn from both positive and hard negative examples dynamically.
RL can be fine-tuned to emphasize relation extraction improvements.

📊 Early Results:
Even with limited training, F1 scores consistently improved, and we saw clear benefits from RL-based optimization. More training = better performance!

🔬 Next Steps:
We’re scaling up experiments with larger models and high-quality data. Stay tuned for updates! Meanwhile, check out one of our experimental models here:
Ihor/Text2Graph-R1-Qwen2.5-0.5b

📔 Learn more details from the blog post: https://medium.com/p/d8b648d9f419

Feel free to share your thoughts and ask questions!

2 replies

·

posted an update 25 days ago

Post

1817

Why choose between strong LLM reasoning and efficient models?

Use DeepSeek to generate high-quality training data, then distil that knowledge into ModernBERT answerdotai/ModernBERT-base for fast, efficient classification.

Blog post: https://danielvanstrien.xyz/posts/2025/deepseek/distil-deepseek-modernbert.html

posted an update 26 days ago

Post

1900

Updated the ColPali Query Generator Space davanstrien/ColPali-Query-Generator to use Qwen/Qwen2.5-VL-7B-Instruct.

Given an input image, it generates several queries along with explanations to justify them. This approach can generate synthetic data for fine-tuning ColPali models.

reacted to fdaudens's post with ❤️ 27 days ago

Post

8556

Yes, DeepSeek R1's release is impressive. But the real story is what happened in just 7 days after:

- Original release: 8 models, 540K downloads. Just the beginning...

- The community turned those open-weight models into +550 NEW models on Hugging Face. Total downloads? 2.5M—nearly 5X the originals.

The reason? DeepSeek models are open-weight, letting anyone build on top of them. Interesting to note that the community focused on quantized versions for better efficiency & accessibility. They want models that use less memory, run faster, and are more energy-efficient.

When you empower builders, innovation explodes. For everyone. 🚀

The most popular community model? @bartowski 's DeepSeek-R1-Distill-Qwen-32B-GGUF version — 1M downloads alone.

4 replies

·

posted an update 27 days ago

Post

2021

🌍 Big step for multilingual AI data!

The Hugging Face community has rated educational content in languages spoken by 1.6 billion people! New additions:
• Japanese
• Italian
• Old High German

Learn more and contribute: https://huggingface.co/blog/davanstrien/fineweb2-community

These ratings can help enhance training data for major world languages.

1 reply

·

reacted to tomaarsen's post with 🔥❤️ about 1 month ago

Post

4611

🏎️ Today I'm introducing a method to train static embedding models that run 100x to 400x faster on CPU than common embedding models, while retaining 85%+ of the quality! Including 2 fully open models: training scripts, datasets, metrics.

We apply our recipe to train 2 Static Embedding models that we release today! We release:
2️⃣ an English Retrieval model and a general-purpose Multilingual similarity model (e.g. classification, clustering, etc.), both Apache 2.0
🧠 my modern training strategy: ideation -> dataset choice -> implementation -> evaluation
📜 my training scripts, using the Sentence Transformers library
📊 my Weights & Biases reports with losses & metrics
📕 my list of 30 training and 13 evaluation datasets

The 2 Static Embedding models have the following properties:
🏎️ Extremely fast, e.g. 107500 sentences per second on a consumer CPU, compared to 270 for 'all-mpnet-base-v2' and 56 for 'gte-large-en-v1.5'
0️⃣ Zero active parameters: No Transformer blocks, no attention, not even a matrix multiplication. Super speed!
📏 No maximum sequence length! Embed texts at any length (note: longer texts may embed worse)
📐 Linear instead of exponential complexity: 2x longer text takes 2x longer, instead of 2.5x or more.
🪆 Matryoshka support: allow you to truncate embeddings with minimal performance loss (e.g. 4x smaller with a 0.56% perf. decrease for English Similarity tasks)

Check out the full blogpost if you'd like to 1) use these lightning-fast models or 2) learn how to train them with consumer-level hardware: https://huggingface.co/blog/static-embeddings

The blogpost contains a lengthy list of possible advancements; I'm very confident that our 2 models are only the tip of the iceberg, and we may be able to get even better performance.

Alternatively, check out the models:
* sentence-transformers/static-retrieval-mrl-en-v1
* sentence-transformers/static-similarity-mrl-multilingual-v1

1 reply

·

reacted to AdinaY's post with 🔥 about 1 month ago

Post

3115

MiniMax, the company behind Hailuo_AI, has joined the open source community by releasing both models and demos of MiniMax-Text-01 & MiniMax-VL-01🔥
- Model
MiniMaxAI/MiniMax-VL-01
MiniMaxAI/MiniMax-Text-01
- Demo
MiniMaxAI/MiniMax-VL-01
MiniMaxAI/MiniMax-Text-01

✨ MiniMax-text-01:
- 456B with 45.9B activated per token
- Combines Lightning Attention, Softmax Attention, and MoE for optimal performance
- Training context up to 1M tokens, inference handles 4M tokens

✨ MiniMax-VL-01:
- ViT-MLP-LLM framework ( non-transformer👀)
- Handles image inputs from 336×336 to 2016×2016
- 694M image-caption pairs + 512B tokens processed across 4 stages

1 reply

·

reacted to AdinaY's post with 🔥 about 1 month ago

Post

3189

MiniCPM-o2.6 🔥 an end-side multimodal LLMs released by OpenBMB from the Chinese community
Model: openbmb/MiniCPM-o-2_6
✨ Real-time English/Chinese conversation, emotion control and ASR/STT
✨ Real-time video/audio understanding
✨ Processes up to 1.8M pixels, leads OCRBench & supports 30+ languages

reacted to their post with 🤗 about 1 month ago

Post

3068

Introducing scandi-fine-web-cleaner davanstrien/scandi-fine-web-cleaner, the first model trained on FineWeb-C community annotations!

FineWeb2 is a massive multilingual dataset for pre-training language models. Like any web-scale dataset, it contains low-quality content. How can we improve it?

Over the past months, an amazing community of 400+ annotators has been labelling content quality (using Argilla) across 23 languages through the FineWeb-C initiative.

Today, I'm happy to share the first classifier trained on this data.

🔍 What we've built:

- A lightweight classifier that efficiently removes low-quality content
- 90%+ precision demonstrated on Danish & Swedish
- Can process the 43M+ documents in Danish FineWeb2 with minimal compute

🌍 Why this matters: The approach can be reproduced for any of the 23 languages in FineWeb-C ( data-is-better-together/fineweb-c). We can improve training data quality at scale without massive compute resources by starting with community annotations and training small, efficient classifiers.

Want to build a classifier for your language? Check out the full blog post with code examples and implementation details: https://danielvanstrien.xyz/posts/2025/FineWeb-c/scandinavian-content-filtering-fineweb.html

1 reply

·

replied to their post about 1 month ago

Model wouldn't be possible without @Lauler @AntonVic @ohallstrom @bjarlestam @menbom @Ekgren @apsod for Swedish and @rasgaard @JakobBlaa @saattrupdan @FrLars21 @markhougaard @KennethEnevoldsen @Apasalic @tqvist @cnila @Soeren-B @KristianL @mathiasn1 @ITK-dev @jannikskytt @AndreasLH @perlausten @sorenmulli @organicoder for Danish!

posted an update about 1 month ago

Post

3068

Introducing scandi-fine-web-cleaner davanstrien/scandi-fine-web-cleaner, the first model trained on FineWeb-C community annotations!

FineWeb2 is a massive multilingual dataset for pre-training language models. Like any web-scale dataset, it contains low-quality content. How can we improve it?

Over the past months, an amazing community of 400+ annotators has been labelling content quality (using Argilla) across 23 languages through the FineWeb-C initiative.

Today, I'm happy to share the first classifier trained on this data.

🔍 What we've built:

- A lightweight classifier that efficiently removes low-quality content
- 90%+ precision demonstrated on Danish & Swedish
- Can process the 43M+ documents in Danish FineWeb2 with minimal compute

🌍 Why this matters: The approach can be reproduced for any of the 23 languages in FineWeb-C ( data-is-better-together/fineweb-c). We can improve training data quality at scale without massive compute resources by starting with community annotations and training small, efficient classifiers.

Want to build a classifier for your language? Check out the full blog post with code examples and implementation details: https://danielvanstrien.xyz/posts/2025/FineWeb-c/scandinavian-content-filtering-fineweb.html

1 reply

·

posted an update about 1 month ago

Post

2251

The data-is-better-together/fineweb-c dataset is growing!

This week a few more languages have got 1,000 annotations for the educational quality of data from HuggingFaceFW/fineweb-2.

Why should you care?

The quality of pre-training data can have a big impact on the performance of downstream language models trained on that data ( HuggingFaceFW/blogpost-fineweb-v1).

Being able to filter by educational quality is on way of improving the quality of the data you use for training an LLM. Very importantly this approach can also reduce the amount of data needed for pertaining.

Why not use an LLM?

LLMs can be used to annotate educational quality for a subset of data. This data can then be used to train a smaller encoder only model to label the full dataset. However, this may not work well for languages outside of english. This is where fineweb-c (community) comes in.

The community is annotating the educational quality of fineweb2 data. Currently 114 languages have some annotations. These annotations will enable a number of things:

- Evaluate whether an LLM can label the educational quality for texts in that language well
- Directly be used for training quality classifiers
- Help discover other rules and huerisitcs for refining fineweb2 further for different languages.

This week the following languages where done:

Swedish thanks to: @Lauler @AntonVic @ohallstrom @bjarlestam @menbom @Ekgren @apsod

Ukrainian thanks to: @hannayukhymenko @robinhad @realPivo @RabotiahovDmytro @reciprocate

Assamese thanks to: @moyoor97 @Arpanjyoti @nawaf-helmi123 @pahigogoi1 @aelhence @kishorekashyap

Want to learn more: https://huggingface.co/blog/davanstrien/fineweb2-community

Contribute yourself here: data-is-better-together/fineweb-c

1 reply

·

reacted to albertvillanova's post with 👀 about 2 months ago

Post

2073

Discover all the improvements in the new version of Lighteval: https://huggingface.co/docs/lighteval/

replied to their post about 2 months ago

There are some already in the Argilla instance!

Croatian: https://data-is-better-together-fineweb-c.hf.space/dataset/f9448cf3-4ede-4329-965f-f9e1d19149a0/annotation-mode

Serbian: https://data-is-better-together-fineweb-c.hf.space/dataset/99b91fe0-4adb-4a84-9eed-770bb12f4104/annotation-mode

You can also join the discussions here: https://huggingface.co/spaces/HuggingFaceFW/discussion :)

replied to their post about 2 months ago

Thanks to the hard work of @ivykopal , the first 1,000 annotations for Slovak have been completed! Make sure to give Ivan a follow :)

Daniel van Strien PRO

AI & ML interests

Recent Activity

Organizations

davanstrien's activity