Tag: #dataset24 results
Nebius releases SWE-rebench-V2, the world's largest open dataset for training coding agents with over 32,000 executable tasks across 20 languages. It includes 120,000+ tasks from real pull requests, filtered for quality using LLM ensembles. A technical report and leaderboard are also available.
A new downloadable corpus for the EU AI Act, featuring 933 legally structured chunks and BGE-M3 embeddings in a single SQLite file. It improves retrieval performance by chunking based on legal provisions rather than arbitrary character windows.
GitHub has released a new repository-level dataset under the CC0-1.0 license. It aggregates multilingual developer content from READMEs, issues, and pull requests to support AI research.
MONET is a new Apache 2.0-licensed dataset containing 104.9 million high-quality image-text pairs. It includes accompanying tools for visualization, retrieval, and T2i model training.
Daimon Infinit, the world's largest multimodal physical world dataset including tactile data, has been jointly released. It attracted participation from Google and numerous universities.
Datamule, Teraflop AI, and Eventual released a 590GB SEC EDGAR dataset on Hugging Face with 8 million filings and 43 billion tokens. It offers free access to parsed financial reports like 10-K and 10-Q, bypassing costly APIs. Data was collected via datamule-python and parsed using selectolax and custom libraries.
GoodPoint curates a 19K ICLR paper dataset annotated with reviewer feedback using author responses, defining effectiveness via validity and author action. It introduces a training recipe with fine-tuning and preference optimization, boosting Qwen3-8B's success rate by 83.7% and achieving SOTA among similar LLMs. Expert human studies confirm higher practical value for authors.
20M+ Indian court cases with structured metadata, citation graphs (followed/distinguished/overruled), Voyage AI embeddings + BM25. Covers Supreme/High Courts/Tribunals; useful for legal NLP, GNNs, RAG eval, Indian lang models. API/bulk JSON/Parquet export, public domain.
Ai2 released MolmoWeb, an open-weight visual web agent in 4B and 8B sizes, with a full training stack and MolmoWebMix dataset featuring 30K human task trajectories across 1,100+ websites. It operates solely from browser screenshots, producing reasoned actions without HTML parsing. This offers an auditable alternative to closed APIs for web automation.
140 million Pokemon players have contributed 30 billion high-precision images for free. This dataset trains robot algorithms to centimeter-level navigation accuracy. Highlights gaming as AI data source.