{"data":{"kind":"file","path":"README.md","version_id":"g70ungc2c9bd5st8bpdm02lf","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1423,"modified_at":"2026-08-29T07:00:22.515000","content_hash":"424d880d0fd9a799fe95a1cd1af6d39d881979f3f6a2efa84193f6e60f68c5a0"},"entries":[],"content":"# mbpp-sandboxed\n\nA code-execution RL environment based on MBPP (Mostly Basic Python Problems), where the agent is given a Python programming task and must write a solution that passes a set of hidden tests.\n\n## What makes this different\n\nThe original MBPP-style environments run the agent's generated code with a raw `exec()` call — meaning untrusted, model-written code executes directly on the host machine with no isolation. This version instead runs every test case inside an **isolated Docker container**:\n\n- No network access (`--network none`)\n- Memory-limited (256MB)\n- CPU-limited (0.5 core)\n- Hard timeout (10s)\n\nThis means broken, malicious, or resource-hungry model-generated code cannot affect the host machine — a meaningful safety improvement for an environment meant to run untrusted model output at scale.\n\n## Task format\n\nThe agent is given a task description plus the tests its code needs to pass, and must respond with a Python solution inside a ```python code block. Score is the fraction of hidden test cases that pass (partial credit, not just pass/fail).\n\n## Dataset\n\nUses the sanitized MBPP dataset (427 hand-verified problems), loaded directly from the original Google Research source to avoid known Hugging Face dataset-conversion issues with mirrored copies.\n\n## Requirements\n\nDocker must be installed and running locally, since test execution happens inside `python:3.11-slim` containers.\n","encoding":"utf-8","truncated":false,"total_bytes":1423},"status":null}