NVIDIA SFT dataset for CUDA code generation
Synthetic dataset of 1k-10k tasks with tests from NVIDIA for training language models to write CUDA kernels.
- Text generation
- English
Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,797 in the catalog so far, with new ones added every hour.
Answered by Gemini 4 · the chat is kept only in this tab
Synthetic dataset of 1k-10k tasks with tests from NVIDIA for training language models to write CUDA kernels.
Dataset for language model SFT fine-tuning to generate coherent reasoning chains in English and Chinese.
An exam-based benchmark for evaluating LLMs in the Japanese financial domain.
Set of 809 questions to test model knowledge in fitness, sports, and health.
Subsample of deduplicated source code identifiers for training generative models.
Multilingual benchmark of original cultural riddles written by native speakers from 62 communities and 45 countries.
Multilingual set of synthetic documents paired with JSON Schema and gold extracted objects for SFT.
Massive corpus of OCR texts from public domain old German newspapers for training language models.
An English benchmark checks whether a coding agent follows rules when working with a repository.
English instruction set for the McEval paper for SFT training and text generation evaluation.
Traditional Chinese medicine clinical cases, classic texts, and QA in Chinese.
A large-scale social media dialogue corpus for training and evaluating language agents with persona clustering.
Large array of cleaned short Python scripts for language model pre-training.
Multi-jurisdiction court decision collection in English and French for legal AI.
About 49,772 traces from qwen3.8-max-preview for SFT and off-policy distillation.
Large open text and QA pair collection based on NIST standards for cybersecurity LLM specialization.
Token count information for all datasets used in pretraining Marin models.
A large-scale set of English synthetic texts generated by an ensemble of language models.
159 legally vetted scenarios from real US courts to test models for deception and fraud propensity.
US companies 10-K annual reports, split into sections and sentences, with sentiment labels.
Structured reviews of Chinese research in materials science, semiconductors, and metallurgy with tables and QA pairs.
A collection of consumer medical search queries released by Google.
A large corpus of 121 thousand instruction-response pairs on pentesting and attacks in English for SFT.
A curated dataset of 3,067 English cybersecurity reasoning chains for SFT training.