Spaces:
Sleeping
Sleeping
Commit History
-
Re-apply training prep work on top of validated Layers 1-5 base
Preserved from the Space-debug branch (backup/space-debug-20260426-0215): - training/prompts.py — class-shaped few-shot + positively-phrased system prompt (replaces the function-only examples that were causing few-shot pollution and the negation-list that was triggering forbidden-import output on small models). - training/train_grpo.py — wires SYSTEM_PROMPT and render_few_shot from training.prompts; build_prompt now injects rendered few-shots before the live task. - src/mutant_hunter/corpus/baselines.py — summarize_module now walks class bodies and emits public methods, properties, and dunders with type annotations and defaults, so the policy can use __contains__ via `in`, __len__ via `len()`, and properties as attribute access without guessing the API. - src/mutant_hunter/corpus/_baselines/*.json — regenerated to reflect the richer module_summary. Mutation scores are unchanged. - scripts/probe_summary_richness.py — local CPU diagnostic for prompt regressions; us…1b83cd4 -
Initial commit: MutantHunter — RL env for mutation-score-rewarded test generation
91487c9
