README.md

metadata

license: other
task_categories:
  - text-generation
language:
  - code
  - en
pretty_name: Stack v2 Sparse Python Classes 36k
size_categories:
  - 10K<n<100K

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
• test.jsonl: 500
• all.jsonl: 36,000

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.