OpenMathInstruct-2: 14M Math Pairs
14 million math problem-solution pairs generated by Llama 3.1 405B for instructional fine-tuning.
- Question answering
- Text generation
- English
Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,788 in the catalog so far, with new ones added every hour.
Answered by Gemini 4 · the chat is kept only in this tab
14 million math problem-solution pairs generated by Llama 3.1 405B for instructional fine-tuning.
Raw data, images, and English text for anomaly detection and segmentation in 3D printing.
A large source code corpus in dozens of programming languages with permissive licenses.
Over a million news articles from 2,000+ sources with topic labeling for text classification.
6.77 billion rated standard games from lichess.org in tabular format for analysis.
500 issue-pull request pairs from Python repositories to evaluate GitHub task resolution via unit tests.
Tabular, textual, and time-series scientific data from Tahoe-100M for analysis and agent queries.
Quantum experiment dataset with detection events for surface code.
Multimodal set of 4,307 characters with AI images, definitions, and pinyin audio for generative and language models.
Continuously updated benchmark of real computational workflows in natural, exact, and engineering sciences.
First dataset with images, text, and time series for semantic communication and 6G tasks.
Expert-level multiple-choice questions in biology, physics, and chemistry to test complex scientific reasoning.
A huge corpus of multimodal instructions (image + text) for full-parameter SFT of vision-language models.
A collection of real tabular datasets from OpenML for comparing classification and regression algorithms.
A modified version of the classic WikiText corpus, where texts are grouped into full pages.
Video and text pairs for training and evaluating video question-answering models.
A stress test for ASR models consisting of 100 audio clips where three people speak different languages simultaneously.
All Hacker News stories, comments, Ask HN, Show HN, jobs, and polls in English.
Curated video fragments and fisheye camera images from the HOT3D benchmark for 3D vision tasks.
Filtered traces of the Claude Code agent CLI without images and service classifier calls.
17.3M images, 24.3M examples, and 88.9M turns for training modern open vision-language models.
Collection of real-world time series from AutoGluon for testing forecasting models.
Precomputed SDXL VAE latents, SigLIP embeddings, and moondream2 captions based on the CommonCatalog CC-BY dataset.
817 questions across 38 categories to test whether a model generates truthful answers.