cactus183/patchbench-dev

1name: patchbench2version: "1.0.0"3description: "Bug-fixing RL environment where agents patch Python code and are graded by actual pytest execution in a sandboxed subprocess. A lightweight SWE-bench analogue fitting in 2 vCPU / 8 GB."4author: "PatchBench Team"5tags:6  - openenv7  - code8  - debugging9  - python10  - pytest11  - real-world12action_space: "Single field patched_code: str containing the full corrected Python source file. Not a diff — a complete replacement."13observation_space: "task_description, buggy_code (current state), failing_tests (human-readable summary of pytest results), step_number, max_steps, reward (from previous step), done, info (diagnostic dict with test counts, regressions, validity)."14tasks:15  - name: easy_0116    difficulty: easy17    max_steps: 518  - name: easy_0219    difficulty: easy20    max_steps: 521  - name: easy_0322    difficulty: easy23    max_steps: 524  - name: medium_0125    difficulty: medium26    max_steps: 827  - name: medium_0228    difficulty: medium29    max_steps: 830  - name: medium_0331    difficulty: medium32    max_steps: 833  - name: hard_0134    difficulty: hard35    max_steps: 1236  - name: hard_0237    difficulty: hard38    max_steps: 1239  - name: hard_0340    difficulty: hard41    max_steps: 1242