Machine Learning Librarian @huggingface 🤗 I like datasets.
Scotland
- Tweets 4,880
- Following 1,532
- Followers 5,786
- Likes 8,169
BioMistral is a new 7B foundation model for medical domains, based on Mistral and further trained PubMed Central.
- top open-source medical Large Language Model (LLM) in its weight class
- Apache License
- includes base models, fine tunes, and quantized versions.
ALT Screenshot of the model card on the Hub.
17
221
1,169
146,071
dolphin-r1: a dataset for training R1-style models
- 800k total samples dataset similar in composition to the data used to train DeepSeek-R1 Distill models.
- 300k from DeepSeek-R1
- 300k from Gemini 2.0 flash thinking
- 200k from Dolphin chat.
ALT Screenshot of the dataset card on the Hugging Face Hub
8
86
510
45,879
DeepSeek-OCR just got @vllm_project support 🚀
Currently processing @natlibscot's 27,915-page handbook collection with one command:
Processing at ~350 images/sec on A100
Using @huggingface Jobs + @astral_sh uv - zero setup batch OCR!
Will share final time + cost when done!
16
40
442
58,113
OpenCodeReasoning: Competitive Coding Dataset
- 735K Python solutions across 28K unique programming problems
- Largest reasoning-based synthetic dataset for code generation
- Generated by NVIDIA's R1 model with full reasoning steps
- Sourced from 10 competitive coding platforms
15
65
373
45,201
Another VLM-based OCR model!
NuMarkdown-8B-Thinking is special, it reasons through documents before converting to markdown
See it analyse this 1878 medical report step-by-step.
Try with @huggingface Jobs + @astral uv - no GPU needed!
6
41
338
23,621
NVIDIA Just Released 8M Sample Open Dataset + OCR Tooling on @huggingface
- 3x larger than v1 (just 2 months ago!)
- Image/video QA, reasoning, multilingual OCR
- Commercial-ready (CC-BY-4.0)
@NVIDIAAI is one of the few major AI labs releasing datasets 🤗
8
48
337
30,729
"UltraTextbooks" on @huggingface is a comprehensive dataset of synthetic and human-written textbooks covering a range of subjects and programming languages.
• 🎓 Over 5.5M examples
• 💾 22.3 GB of educational content
• 🌐 From math to science, and code
ALT Screenshot of the dataset card for the UltraTextBooks dataset on the Hugging Face Hub
4
49
309
42,095
Need blazing-fast classifier inference with minimal code?
ModernBERT now runs on @vllm_project — fast enough to process 200K+ arXiv papers in minutes.
It makes running any of the 100s of ModernBERT models on the @huggingface Hub even quicker.
Guide 👇
danielvanstrien.xyz/posts/20…
5
59
300
29,600
AM-DeepSeek-R1-Distilled-1.4M: Massive reasoning dataset for LLM training
- 1.4M high-quality reasoning problems with verified solutions
- 900K entries distilled from DeepSeek-R1-671B
- Covers math, code, and complex reasoning tasks
- Bilingual (Chinese/English)
10
58
293
24,669
RLVR: Breaking RL Beyond Math & Coding Boundaries
- Extends reinforcement learning to 48 diverse academic fields
- 638K expert-crafted QA pairs in medicine, law, economics & more
- Proves RLVR effective for complex reasoning across all disciplines
- Apache 2.0 licensed
11
37
250
29,374
Ultra-FineWeb: A cleaner 1.1T-token foundation for better LLMs
– 1T English + 120B Chinese, filtered for quality
– +3.6 pts on MMLU, +3.7 on CMMLU vs FineWeb
– Verification cut from 1200 GPUh → 110
– FastText classifier: 6,000 GPUh → 1,000 CPUh
9
45
243
44,297
🤗 Ridiculously excited that today is my first-day working at @huggingface! I genuinely believe Hugging Face has done so much to democratise machine learning and I'm looking forward to contributing to this mission every day!
15
11
226
43,830
Excited to announce #BigLAM, a @BigscienceW x @huggingface Hackathon focused on making library, archive, and museum datasets related to machine learning more accessible and discoverable github.com/bigscience-worksh…
5
85
213
Want to get started using #machinelearning based computer vision methods for humanities research/#DH/#GLAM applications? 💡
ALT Screenshot with the following text: Computer Vision for the Humanities: An Introduction to Deep Learning for Image Classification (Part 1) Daniel van Strien, Kaspar Beelen, Melvin Wevers, Thomas Smits and Katherine McDonough This is the first of a two-part lesson introducing deep learning based computer vision methods for humanities research. Using a dataset of historical newspaper advertisements and the fastai Python library, the lesson walks through the pipeline of training a computer vision model to perform image classification.
2
68
166
ColPali is revolutionizing multimodal retrieval, but could it be even more effective with domain-specific fine-tuning?
Check out my latest blog post, where I guide you through creating a ColPali fine-tuning dataset using @Alibaba_Qwen's Qwen2-VL-7B-Instruct model to generate queries.
8
30
197
31,936
Visualizing the embedding space for models from @huggingface Hub using @nomic_ai's Atlas tool and @JinaAI_ embeddings. 🧵
Check out the overall embedding map! Clear clusters emerge, along with intriguing outliers. Let's dive into this exploration!
ALT A screenshot of a 'map' showing various coloured clusters representing model cards on the Hugging Face Hub.
5
28
173
47,495
KodCode: The Largest Verified Synthetic Coding Dataset
- 447K question-solution-test triplets with verifiable correctness
- 12 diverse subsets
- 10-trial verification system for solution robustness
- Each problem includes automated test case verification
12
38
167
21,250
The Mistral-7B-v0.1 model by @MistralAI is the foundation for at least 178 models available on the @huggingface Hub.
These fine-tuned models have collectively been downloaded over 300,000 times.
Strong, openly shared base models genuinely impact open-source machine learning.
ALT Screenshot of the following text: Models fine-tuned from mistralai/Mistral-7B-v0.1 mistralai/Mistral-7B-v0.1 has 178 children mistralai/Mistral-7B-v0.1's children have been downloaded 328,312 times
5
27
162
39,340
MegaMath: Pushing the Limits of Open Math Corpora
- 213.1B tokens across specialized math content domains
- Largest math pretraining dataset, surpassing DeepSeekMath by 30%
- Rigorous deduplication & fontset filtering for quality
- Delivers 15-20% boost on math benchmarks
8
33
167
25,067
Access over 32 Billion tokens of public domain historical newspaper data in 10 European languages. It's available on the @huggingface Hub, courtesy of biglam/europeana_newspapers. Documentation is a work in progress, but you can already explore this collection. ✨🗞️🌍📰🔎
8
35
154
24,411
This is a work-in-progress Space where you can interact with the children of Mistral-7B-v0.1.
If anyone has excellent network visualization skills, I would be thrilled to provide the data so that you can enhance its aesthetic appeal!
huggingface.co/spaces/davans…
ALT A network diagram showing a bunch of nodes and links
7
31
147
38,787
BERTopic now has @huggingface hub integration! We're excited to see which models you share on the Hub 🤗
🤔 Want some inspiration? Check out this notebook which shows you how to train a topic model on Transformers @github issues and shares it on the Hub
colab.research.google.com/#f…
1
25
143
26,753
Introducing FineWeb-C 🌐🎓, a community-built dataset for improving language models in ALL languages.
Inspired by FineWeb-Edu the community is labelling the educational quality of texts for many languages.
318 annotators, 32K+ annotations, 12 languages - and growing! 🌍
ALT Text saying FineWeb-c Educational content in many languages, labelled by the community
2
28
145
18,502
DeepReviewer-13K: A dataset for training LLMs for academic paper review
- 13,378 high-quality structured reviews
- 33.24% accept rate
- Multi-stage reasoning with novelty and reliability analysis
- Available in 6 languages
- Custom license preventing use in formal reviews
5
34
135
16,692
Glaive Reasoning Dataset: Filling the Non-Technical Reasoning Gap
- Addresses lack of large, general reasoning datasets
- 177GB of reasoning traces beyond math/code
- Creative writing & conversation scenarios
- Apache-2.0 licensed for open research
11
19
136
17,376
GeneralThought-195K: Diverse Reasoning Dataset
- 195K reasoning traces from 7+ models
- Expanded beyond math to sciences, humanities & conversations
- Full reasoning traces with verification scores
- MIT licensed with community contributions
3
26
137
19,680
Big-Math: Big-Math: Massive Math Dataset for RL Training
- 10x larger than GSM8k/MATH
- 3 core properties: uniquely verifiable, open-ended, closed-form
- Human-validated 90%+ precision filters
- Difficulty metrics for curriculum learning
2
25
130
19,215
WILDCHAT-50M: The largest open chat dataset
- 125M+ chat transcripts
- 1M+ conversations per model
- Built on WildChat
- Used to create RE-WILD SFT mix
Outperforms existing benchmarks with 40% less data
ALT Screenshot of the wildChat collection on Hugging Face
1
21
129
18,008
Open models (GLM-4.6, Kimi K2, DeepSeek, etc.) + @opencode + @huggingface Inference Providers = automated GitHub code reviews.
Tested on real repos. `/oc fix this` → bot creates PR.
Works great, costs pennies. 5 minutes to set up!
Guide: huggingface.co/docs/inferenc…
8
20
129
19,118
Llama-Nemotron-Post-Training-Dataset-v1: Massive Model Training Dataset
- 30M+ examples: 19.8M math, 9.6M code + science, instruction & safety data
- Powers NVIDIA's Llama-3.3-Nemotron-Super-49B & 3.1-Nemotron-Nano-8B
- Responses generated by 9 leading foundation models
1
19
122
4,721
AutoMathText is a 200 GB dataset of mathematical texts, autonomously labelled for quality and relevance and available on @huggingface Hub
📚 Sourced from websites, arXiv, and GitHub
🤖 Scored by Qwen-72B
Potentially very powerful for making open LLMs better at maths
ALT Screenshot of a dataset preview of the AutoMathText dataset on the Hugging Face Hub
24
117
14,106
Recently, @ai4privacy released "the world's largest open-source privacy dataset" on the @huggingface Hub. Here are its features:
🏷️ 54 PII classes
💼 229 use cases
🌐 Multilingual: EN, FR, DE, IT
ALT Screenshot of a Hugging Face dataset (https://huggingface.co/datasets/ai4privacy/pii-masking-200k). Shows some metadata and the first few rows of the dataset. The preview of the dataset shows columns related to masked and unmasked text and span labels.
1
16
115
38,003
AceCode-89K: First automated test synthesis pipeline
- 89K coding problems with reliable test cases
- GPT-4o-mini generates 16 test cases per problem
- Pass rates used as verifiable rewards
- Enables RM training & RL for coding models
8
23
119
19,357
Great @huggingface blog post from @ihor_step on replicating R1 for text-to-graph tasks. It's great to see GRPO adapted to reward an LLM for doing a novel task well.
🔗 huggingface.co/blog/Ihor/rep…
34
120
6,035
Arboretum is the largest public biodiversity dataset with 134.6M curated images from iNaturalist. It features image-language paired data for training multimodal AI models to support biodiversity and agriculture research.
@huggingface dataset: huggingface.co/datasets/Chih…
ALT Screenshot of a species images in the dataset.
1
25
114
16,801
Synthetic data is going to be massively important in 2024, so we have recently launched a new tag on the @huggingface Hub to facilitate the discovery and sharing of synthetic datasets. To add this tag to your dataset card metadata, simply include the `synthetic` tag.
2
17
107
23,915
Yesterday, I shared a blog post on generating data for fine-tuning ColPali using @Alibaba_Qwen's Qwen2-VL-7B-Instruct.
To simplify testing this approach, I created a @gradio Space that lets you generate queries from an input document page image.
huggingface.co/spaces/davans…
4
16
115
19,412
LEXam: Legal Reasoning Benchmark
- 4,586 exam questions from Swiss, EU & international law
- 3 configurations: standard MCQs, perturbed MCQs & open questions
- Evaluated 20+ SoTA LLMs with expert verification
- Rich metadata across jurisdictions & legal domains
6
21
115
17,555
OpenThoughts2-1M: Million-sample dataset powering SOTA reasoning models
- Systematically curated from 26 data generation approaches
- Powers models outperforming DeepSeek-R1-32B on math benchmarks
- Enables 76.7% on AIME24 & 90.8% on MATH500 with just SFT
5
28
114
6,406
WildChat from @allen_ai consists of 1 million exchanges between ChatGPT and human users, complemented by demographic details like location, IP addresses, and request headers. Check it out at: huggingface.co/datasets/alle…
6
21
111
30,587
465 people. 122 languages. 58,185 annotations!
FineWeb-C v1 is complete! Communities worldwide have built their own educational quality datasets, proving that we don't need to wait for big tech to support languages.
Huge thanks to all who contributed!
huggingface.co/blog/davanstr…
4
28
108
20,113
DeepCoder: Comprehensive Programming Dataset
- 24K+ verified problems with 5+ test cases each
- Sources include TACO, PrimeIntellect, LiveCodeBench, Codeforces
- Temporal split: train on pre-Aug 2024, test on newer problems
- Powers SOTA open-source 14B DeepCoder model
3
18
108
4,533
OCR is one of AI's oldest challenges (first systems: early 1900s!)
Modern vision-language models have transformed what's possible: handwriting, 100+ languages, math formulas, tables, signature extraction...
New @huggingface guide on OCR
huggingface.co/blog/ocr-open…
1
21
112
5,339
Reasoning datasets dominate @huggingface trending datasets, but focus mainly on code/math.
We launched a competition with @bespokelabsai & @togethercompute to diversify this. Create a PoC reasoning dataset and win prizes to help scale it!
2
29
108
22,176
REALM: Real-World Application of Large Language Models Dataset
- 93k+ documented LLM use cases from 2020-2024
- Reddit posts + news articles with full content
- Dual categorization: AI Use Taxonomy & O*NET
- MIT-licensed with interactive exploration dashboard
1
10
107
16,724
You can now generate synthetic data using @OpenAI's GPT OSS models on @huggingface Jobs!
One command, no setup:
hf jobs uv run --flavor l4x4 [script-url] \
--input-dataset your/dataset \
--output-dataset your/output
Works on L4 GPUs ⚡
huggingface.co/datasets/uv-s…
3
23
104
16,317
Can we combine the power of the @huggingface datasets hub and @LabelStudioHQ to create a workflow that allows you to move quickly between annotating and model training?
new blog post exploring this 👉🏻 danielvanstrien.xyz/huggingf… 🧵
3
30
100
SCP-116K: A massive scientific problem-solving dataset
- 116K+ high-quality problem-solution pairs
- Covers physics, chemistry & biology
- University to PhD-level content
- Model-generated solutions included
1
26
98
14,854
Reasoning models aren't just for math and science!
Using @Alibaba_Qwen's QwQ-32B + @bespokelabsai Curator to create reasoning datasets for structured data extraction from model cards.
First in a series on using GRPO for increasingly weird things...
danielvanstrien.xyz/posts/20…
4
16
100
6,368
Do you need a dataset to train a custom sentence transformer model? I've created a pipeline for using an LLM to create a synthetic dataset you can directly use for fine-tuning/training a Setence Transformers model.
*Link in next tweet
3
19
100
14,894
Want to convert a bunch of PDFs into a dataset of single-page images? There's a @huggingface Space for that!
huggingface.co/spaces/Datase…
ALT Illustration showing a single PDF document being split into separate pages
5
20
100
23,167
WebInstruct-CFT: Teaching LLMs to critique
- 600K instruction-critique pairs
- 65% math, plus business & sciences
- GPT-4 generated detailed critiques
- 3 sizes: 4K/50K/600K examples
ALT Screenshot of the dataset viewer on the Hugging Face Hub.
2
13
98
4,527
🚀 Introducing Hugging Face Similar: a Chrome extension to find relevant datasets!
✨ Adds a "Similar Datasets" section to @huggingface dataset pages
🔍 Recommendations based on dataset READMEs
🏗️ Powered by @trychroma and @SnowflakeDB
Get it now!
chromewebstore.google.com/de…
3
20
92
26,787
Embedding @huggingface Hub dataset cards with existing models was challenging due to the limited context window. Now, with the new long context embedding models from @JinaAI_ and @nomic_ai Atlas, we can visualize the full context of these cards.
🔗: atlas.nomic.ai/map/393ab060-…
ALT Screenshot of a embedding map consisting of different clusters of dataset cards.
4
24
90
23,743
Hot take: shipping BERT-sized models in 2025 will benefit far more people than sharing an LLM overfitted to some saturated leaderboards 🙊
We're already seeing ModernBERT finetunes on the @huggingface Hub. My guess is we'll see hundreds of these by the end of 2025.
ALT Screenshot of a model tree of ModernBERT-base fine tunes
4
6
95
6,438
ColPali is an exciting new approach to multimodal document retrieval, but some doubt its practical use with existing vector DBs.
It turns out it's super easy to use @qdrant_engine to index and search ColPali embeddings efficiently.
Blog post here: danielvanstrien.xyz/posts/po…
ALT Screenshot of a search for "top secret" in a function and an image result below which has a page from a document with UFOs.
5
22
96
11,785
WildDoc: Real-World Document Understanding Benchmark
- 12K+ document images captured in natural environments
- Each document photographed in 4 different conditions
- Tests 5 real-world factors
- Reveals 35.3% performance drop in leading MLLMs
7
11
94
16,202
DoTS.ocr from @xiaohongshu just got native @vllm_project support!
I built a UV script so you can run SOTA multilingual OCR in seconds with zero setup using @huggingface Jobs
Tested on 1800s library cards - works great ✨
2
11
94
37,531
Academic Chains: Scientific Thinking for LLMs
- Distils reasoning from papers in biology & economics
- Captures researcher intuition & exploration
- Fine-tuning shows 7.2% improvement on MMLU-Pro Economics
- @huggingface /@bespokelabsai reasoning datasets competition submission
7
16
92
15,398
Is your desktop full of random screenshots? I just wrote a blog post on using @MistralAI's Pixtral (via @LMStudioAI) to automatically organize screenshots—no cloud APIs needed, so you can keep the weird memes you collect private!
Post here: danielvanstrien.xyz/posts/20…. More info 🧵
4
12
91
16,719
APIGen from @sfresearch is a function-calling dataset created by APIGen, "an automated data generation pipeline designed to produce verifiable high-quality datasets for function-calling applications." 60,000 rows, 3,673 APIs, 21 categories
huggingface.co/datasets/Sale…
ALT Image showing a workflow for generating data synthetically
18
86
17,765
I just processed 1000s of prompts using Qwen3-235B-A22B-Instruct-2507 across 4 GPUs!
How? Everyone plays their part:
@astral_sh UV handles dependencies
@huggingface Jobs handles GPUs
@Alibaba_Qwen handles the model
@vllm_project handles inference
One command. Zero complexity!
2
13
86
17,213
Awesome to see @hackernoon share a dataset of tech news on the @huggingface Hub.
📊 6.9 Millions Rows
📜 MIT licence
🔍 Datasets server preview
ALT Screenshot of the huggingface server preview for the dataset
5
21
79
30,848
Finally documented the Beyond Words dataset from @LC_Labs/@lee_bcg for BigLAM @huggingface org!
- 3.5K annotated historical newspaper pages
- Bounding boxes + category labels
- Photos, ads, headlines, cartoons & more
Also trained some YOLO 11 models with the dataset for fun!
2
15
84
15,621
II-Thought RL v0: First large-scale multi-domain RL dataset
- 960K+ high-quality question-answer pairs
- 735K math problems from multiple sources
- 98K code problems with execution validation
- 74K science & specialized domain samples
3
16
88
4,789
AgentTrek: Web Agent Training Dataset
- 52K annotated dialogue turns
- Web browsing & shopping tasks
- User-agent interaction flows
- Real-world web task simulations
1
10
86
17,554
📁✨Meet Corpus Creator! This @Gradio app turns your local files into a chunked @huggingface dataset via @llama_index. Perfect for building datasets for synthetic data pipelines, annotation, and beyond. Try it here: huggingface.co/spaces/davans…
1
20
83
23,114
How do you make 1M+ @HuggingFace models & datasets more discoverable? 🤔
I fine-tuned SmolLM2-360M to generate one-line summaries from a README. Its own self-description?
"A model for generating concise summaries of model & dataset cards from the Hugging Face Hub"
4
11
85
16,242
EuroSpeech: Massive Multilingual Parliamentary Speech Corpus
- 78,100+ hours across 22 European languages
- 50,500+ hours of quality-filtered data (CER < 20%)
- Robust alignment algorithm for non-verbatim texts
- Dramatically expands resources for 19+ languages
2
9
86
6,911
Get ready for 2024: The Year of Open Machine Learning Datasets! 🌟
Here's a mega thread for all the datasets I'm showcasing this year. 🧵
ALT An illustration in soft pastel colors featuring a whimsical robot cat seated in a cozy armchair. The robot cat, combining feline features with robotic elements, has a friendly and playful expression. It is attentively reading a book titled '2024: The Year of Open Machine Learning Datasets'. Surrounding the robot cat are various open books and digital screens, each displaying colorful graphs and data from different machine learning datasets. The room is decorated in a storybook fashion, complete with fluffy cushions, plush furnishings, and gentle sunlight streaming through a window, creating a warm and inviting atmosphere.
3
10
78
29,326
Excited to introduce Synthetic Data Workshop, a @HuggingFace Space that aims to make creating synthetic datasets easier!
✅ Pre-configured environment
✅ Ready-to-use notebooks
✅ No local GPU needed
huggingface.co/spaces/davans…
ALT Screenshot of the landing page for the Space
4
16
83
5,629
SwallowCode: LLM-Rewritten Python Dataset
- 16.1B tokens from The Stack v2
- Filtered by syntax + pylint (≥7.0)
- Rewritten twice via Llama-3.3
- +17.0 pass@1 (HumanEval)
3
18
78
19,618
Multimodal ArXiv is a Dataset for Improving Scientific Comprehension of Large Vision-Language Models
🖼️ figure-caption dataset
📸 6.4M images and 3.9M captions
📄 from 572K ArXiv papers spanning various scientific domains.
🤗 Available on @huggingface Hub.
ALT Screenshot of MMInstruction/ArxivCap dataset on the Hugging Face Hub
3
14
83
20,745
NVIDIA's ClimbLab: Setting a New Standard for Pretraining
- 1.2 trillion tokens in 20 semantic clusters
- Two-classifier system removes low-quality content
- Demonstrates superior scaling properties in 1B models
- CC BY-NC 4.0 licensed for research community
3
18
76
14,739
Fine Reasoning Questions: Expanding Web Text to Reasoning Tasks
- 144 complex reasoning questions from diverse web content
- Both text-dependent and independent formats beyond math & science
- Aims to show how to transform "in the wild"web content into reasoning questions
1
21
80
13,237
MIRACLRetrieval: Massive Multilingual Search Dataset
– 18 languages from 10 language families
– 78K queries with 726K+ relevance judgments
– 106M+ unique Wikipedia documents
– Expert annotations by native speakers
4
8
74
15,541
Turn any HF dataset into an interactive embedding visualization - now with @huggingface Jobs support!
One command → GPU processing → deployed Space
Built on @Apple's Embedding Atlas library ❤️
2
11
76
7,299
GSM8K-Platinum: Enhanced Math Benchmark
- 1,209 math problems after removing 110 ambiguous questions
- Identified 8.3% of problems had errors in original benchmark
- Prevents false "benchmark saturation"
- Even 96% accurate models make elementary reasoning errors
3
13
76
15,281
Thanks to @tomaarsen's work in the latest Sentence Transformers release, training custom models is easier than ever.
With improved training support and synthetic data for fine-tuning, you can build a model in less than a day.
Example here👇: huggingface.co/collections/d…
ALT Screenshot of a code model and some before after metrics
1
8
76
35,862
TimeTravel: A Comprehensive Benchmark to Evaluate LMMs on Historical and Cultural Artifacts
- 10K+ expert-verified artifacts
- 266 cultures across 10 regions
- Full metadata & image pairs
9
10
76
14,980
NESTFUL: A benchmark for nested API calls from @IBMResearch
- 1.8K+ executable function sequences
- Tests math reasoning & coding tools
- Evaluates variable handling & chaining
- Shows gaps in current LLM capabilities
Apache 2.0 licensed & fully reproducible
ALT Screenshot of the dataset viewer on the Hugging Face Hub
6
10
69
35,285
Super excited to have a guest post on the @huggingface blog showing how you can use 🤗Datasets and Sentence Transformers to create an image search app for 19th Century book images huggingface.co/blog/image-se… 🧵
1
20
76
🗞️ Finally documented this massive multilingual newspaper dataset we curated for BigLAM — built from the amazing @Europeanaeu Newspapers collection.
- 32B tokens
- 12 languages
- OCR scores + metadata
Find it on @huggingface
3
20
76
16,489
"Can language models (LLMs) understand protein sequences like natural language?" ProteinLMDataset is a dataset with 17.46 billion tokens for pretraining and 893,000 instructions for supervised fine-tuning (SFT) that helps explore these questions.
huggingface.co/datasets/tsyn….
2
13
74
17,810
Created an "Awesome Synthetic Datasets" list in my ongoing quest to learn more about building synthetic datasets using large language models. Currently includes important tools, datasets, and papers.
Check it out here: github.com/davanstrien/aweso…
1
15
72
5,976
CuratedThoughts: Clean Math Data for RL
- Fixes critical flaws in math reasoning datasets
- Removes 5-25% of problematic examples unsuitable for RL
- Prevents models from learning invalid reasoning paths
Enables reliable reward verification for GRPO training
2
18
73
33,979
RLAIF-V-Dataset is a large multimodal feedback dataset featuring images + questions alongside chosen and rejected responses.
Available on the @huggingface Hub: huggingface.co/datasets/Haoy…
21
76
18,346
An intriguing new 7B merge model has just been shared on the @huggingface Hub. The model is an experimental combination of Zephyr 7B Beta and Notus 7B v1. Cool to see smaller merges also being created!
Check it out here: huggingface.co/mergedlm/zeph…
7
7
72
9,074
Great to see @Alibaba_Qwen sharing datasets on @huggingface!
Many/most bigger labs are not doing this!
2
8
73
3,068
I was excited to see that @huggingface datasets recently added support for image features 📸. I wrote a little blog post on how you can use datasets + faiss + Sentence Transformers to create an image search application for searching bl book images 👉🏻 danielvanstrien.xyz/metadata…
11
73
@nanonets just shipped Nanonets-OCR2: new 3B VLM for OCR!
LaTeX equations, tables, handwriting, charts, multilingual - it does it all!
You can try it against your data with one command via @huggingface Jobs - no local GPU needed!
The HF Jobs command/output from the model 👇
6
14
73
24,427
Introducing Haiku DPO: a synthetically generated Direct Preference Optimization dataset. It consists of user prompts for a haiku and generations scored for "correctness".
You can directly use this dataset for DPO, but it may also be interesting for exploring other questions 🧵
2
15
68
10,652
Already doing a yolo fine-tune run of Qwen3-VL-4B
Using a @huggingface Jobs script so just have to switch out a model ID!
2
9
70
4,024
HF Jobs just launched! 🚀
One command VLM based OCR with uv Scripts:
hf jobs uv run [script] ufo-images ufo-text
Classified UFO docs → clean markdown. Zero setup!
Try it → huggingface.co/uv-scripts
3
5
69
17,066
Over 1.5M models on @huggingface…
How do you pick the right one for your needs?
🔍 Try this semantic search prototype with size filters (0-1B to 70B+):
🔗 huggingface.co/spaces/librar…
5
12
66
14,522
Great to see community building on top of each other's work for datasets. @argilla_io took the excellent OpenHermes-2.5 dataset from @teknium, generated new responses using a @NousResearch model and used Distilabel + PairRM from @billyuchenlin et al. to create a DPO version.
ALT Screenshot of trending datasets with two datasets highlighted with a yellow circle indicating one dataset was used to help create another dataset.
1
11
62
16,271
Instruction pre-training is a new approach that enhances LLM pretraining by using instruction-response pairs from an instruction synthesizer instead of raw data.
Explore this method in this @gradio Space: huggingface.co/spaces/davans…
ALT A workflow diagram showing the use of a instruction synethizer
ALT An example screenshot of the gradio demo.
1
20
67
14,185
Do you want to keep track of new models trained on a dataset from the @huggingface Hub? The Dataset-to-Model Monitor will notify you whenever a new model is released that uses a dataset you are tracking. huggingface.co/spaces/librar…
ALT A screenshot of a notification about a new model. The notification is a Hugging Face discussion which tags people tracking a dataset.
ALT A screenshot of the app which shows some describing the app. It also shows a text box which allows you to input the ID for a dataset to track. Other buttons allow you to list datasets a user is tracking and unsubscribe from alerts.
3
11
62
28,637