{"data":{"kind":"file","path":"README.md","version_id":"zeay715b67nnahz5v8hyzruu","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1249,"modified_at":"2026-08-05T03:48:33.157000","content_hash":"dba5d2f9bf230f35efd9335626e06d4e7ff49ca315a4455dcdc25d2fdb16a739"},"entries":[],"content":"# Mini SWE-Bench\n\nA verifiers environment that evaluates LLM bug-fixing capabilities on real-world Python bugs.\n\n## Overview\n\nThis environment presents models with buggy code snippets from real open-source Python projects. The model must analyze the bug description and produce a correct fix. Tasks cover common bug categories:\n\n- **Off-by-one errors** — incorrect loop bounds, fencepost errors\n- **Type errors** — wrong type handling, missing conversions\n- **Logic errors** — incorrect conditions, wrong operators\n- **API misuse** — incorrect function signatures, wrong argument order\n- **Edge cases** — empty inputs, None values, boundary conditions\n- **String/encoding bugs** — incorrect parsing, encoding issues\n- **Concurrency bugs** — race conditions, missing locks\n\n## Evaluation\n\nEach task is scored on:\n1. **Fix correctness** (weight 0.5) — does the fix resolve the described bug?\n2. **Code quality** (weight 0.3) — is the fix clean and idiomatic?\n3. **Preservation** (weight 0.2) — does the fix avoid breaking existing functionality?\n\n## Usage\n\n```bash\nprime eval run davida/mini-swe-bench -m meta-llama/Llama-3.2-1B-Instruct\n```\n\n## Tags\n\n`swe-bench`, `bug-fix`, `code-generation`, `software-engineering`, `debugging`\n","encoding":"utf-8","truncated":false,"total_bytes":1249},"status":null}