README.md

metadata

license: other
task_categories:
  - text-generation
language:
  - code
  - en
pretty_name: Stack v2 Sparse Python Classes 75kplus
size_categories:
  - 10K<n<100K

Stack v2 Sparse Python Classes 75kplus

This is a frozen snapshot with 75829 samples for Diffusion + Autoregressive hybrid code generation experiments.

Splits

• train.jsonl: 74829
• val.jsonl: 500
• test.jsonl: 500
• all.jsonl: 75829

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.

This snapshot freezes the first 75829 JSONL rows from the live local dataset at packaging time, then shuffles with seed 42 before splitting.

Record Format

Each JSONL row is one Python class sample. Important fields include:

• prompt: natural-language class implementation prompt
• skeleton: class/method signatures and docstrings with <|body_i|> slots
• bodies: list of method bodies without docstrings
• bodies_text: body slots wrapped by <|body_start_i|> and <|end_body_i|>
• full_text: skeleton plus body slots
• solution: reconstructed class code
• source_repo, source_path, revision_id, blob_id, detected_licenses: source metadata

Filters

• 2 to 6 methods per class
• every method has a non-empty docstring
• every method body has 3 to 30 non-empty lines
• reconstructed class parses as Python AST
• tests/docs/examples/vendor/generated files are excluded by metadata/path filters
• simple ClassEval/HumanEval contamination filters are applied

Strict pyflakes is not used as a hard filter because isolated extracted classes often depend on module-level imports, constants, parent classes, or helper functions.