mlabonne/llm-datasets

Organizes post-training datasets across specialized categories (instruction, math, code, reasoning) with quality criteria emphasizing accuracy, diversity, and complexity—evaluated through manual review, heuristics, and judge LLM scoring. Covers both human-curated and synthetically-generated datasets from academic and industry sources, with focus on permissive licensing. Integrates primarily with Hugging Face Hub for dataset hosting and targets the SFT/fine-tuning phase of LLM development pipelines.

4,319 stars. Actively maintained with 2 commits in the last 30 days.