Datasets:

rmems
/

python-function-repair-20261006

pretty_name: Python Function Repair Records (2026-10-06)
license: apache-2.0
language:
  - en
tags:
  - synthetic-data
  - code-repair
  - software-engineering
  - deterministic-execution
  - procedural
  - curation
  - provenance
  - research-snapshot
configs:
  - config_name: viewer
    data_files:
      - split: train
        path: data/viewer/records.parquet
dataset_info:
  - config_name: viewer
    features:
      - name: source_file
        dtype: string
      - name: source_line
        dtype: int64
      - name: record_json
        dtype: string
    splits:
      - name: train
        num_examples: 132
        num_bytes: 3542499

Python Function Repair Records (2026-10-06, deterministic)

Generated: 2026-10-06 (run cr-20261006, produced_at 2026-10-06T00:00:00Z) · Published to the Hub: 2026-10-07

Rights & intended use: public research snapshot, not training data. Deterministic, project-owned generation over real MIT-licensed upstream programs (TheAlgorithms/Python @ 2067ce6dfb3, MIT — see provenance.json): intended_use: training_candidate, project_training_policy: allowed per the factory registry — but this raw snapshot is not training-ready (release-status.json: release_stage: raw_uncurated_public, training_ready: false). Machine-readable record: provenance.json. License: Apache-2.0 for generated content; upstream programs remain MIT by their authors. There is no rights.json: the repo's rights schema (v0.1.0) only models hosted-frontier provider routes, so procedural rights are declared here and in provenance.json instead.

Verification: every candidate carries its execution evidence, but independent replay covered only part of the batch: evidence/REPLAY.json reports 97 of 132 candidates replayed and passed, 0 failed, and 35 not replayed (all 22 rejected rows plus 13 accepted rows whose hidden check was unavailable); its overall status: passed applies to the replayed subset. The 132 rows mix 110 accepted and 22 rejected outcomes, with 115 validated and 17 provisional oracle statuses (evidence/RUN.json); per-row validation.status is unvalidated. evidence/execution-log.jsonl is harness telemetry, not training rows, and is excluded from the viewer projection.

132 candidates (110 accepted, 22 rejected) from 150 requested draws: 18 draws were skipped as duplicate mutants; separately, 4 of the 197 catalog programs have no mutation sites and were excluded from the draw pool:

file	records	sha256
`data/raw/candidates.jsonl`	132	`743327ef7a49`…

Viewer projection: data/viewer/records.parquet (132 rows, lossless source_file / source_line / record_json; Data Studio viewer config).

Reproduce

From the factory repo: python3 pipelines/code_repair_cli.py generate --catalog catalogs/python-repair-v1 --seed 20261006 --count 150 --out <run-dir> then replay --run <run-dir> --catalog catalogs/python-repair-v1 --out <replay-dir>.

Python Function Repair Records (2026-10-06, deterministic)

Contents

Reproduce