poolside-laguna-hackathon/patchrecoverygym-laguna
收藏资源简介:
PatchRecoveryGym for Laguna 是一个可复现的评估和强化学习环境,旨在测试编码代理是否能够从错误的首次尝试中恢复——这是一个实际且未被充分测量的代理编码弱点。该环境专为Poolside Laguna XS.2设计,针对依赖迁移修复任务(例如Pydantic v1到v2、Ruff严格性)。每个任务包括任务描述、一个错误的首次尝试以及该尝试的失败测试输出,模型必须生成一个修正后的最小补丁。奖励基于隐藏测试是否通过(二进制,不可作弊)。其重要性在于:代理编码器通常能编写看似合理的修复,但很少能从自己的错误尝试和错误消息中纠正方向;大多数基准测试仅测试从零开始解决,而此数据集则专注于隔离恢复能力。
PatchRecoveryGym for Laguna is a reproducible eval + RL environment that tests whether a coding agent can recover from a wrong first attempt — a real, under-measured agentic-coding weakness. Built for Poolside Laguna XS.2 on dependency-migration repair tasks (e.g., Pydantic v1→v2, Ruff strictness). Each task provides the task description, a wrong first attempt, and that attempts failing test output, and the model must produce a corrected minimal patch. Reward is based on whether the hidden tests pass (binary, ungameable). Why it matters: agentic coders write plausible-looking fixes but rarely course-correct from their own bad attempt + an error message. Most benchmarks test solve-from-scratch; this one isolates recovery.




