hybrid-diff-ar/stack-v2-sparse-classes-36k

Stack v2 Sparse Python Classes 36k This is a 36,000-sample snapshot for Diffusion + Autoregressive hybrid code generation experiments. Source The data is extracted from bigcode/the-stack-v2-dedup, Python subset. The extraction uses Stack v2 metadata as source of truth, groups candidates by repo_name + revision_id, fetches files with git partial fetch + sparse checkout, then applies AST-level class filters. Splits train.jsonl: 35,000 val.jsonl: 500… See the full description on the dataset page: https://huggingface.co/datasets/hybrid-diff-ar/stack-v2-sparse-classes-36k.