Zhiyuan Hu

Mixed-reasoning full SFT snapshot on Genshin + Cyberpunk 2077 + Spider-Man 2 (0916 polish).

DeepSeek-V4.1 reward-1 trajectories for tasks on which the Nemotron student did not obtain reward 1. This release was rebuilt from the complete reward-1 audit under v56-shell-quote-aware-package-probes+v57-target-source-recheck. 每个目录内的文件名都直接说明：是一题一条还是全部 trials、是否为 audit、thinking/action 是否干净。 本版重新检查了每条候选轨迹的实际 tool action。查询、下载或读取当前任务项目的已发布源码时， 即使 verifier 给出 reward=1，也不会进入上述训练文件。审计隔离清单位于 data/03allreward1trials/targetpublishedsourcelookupquarantineaudit.jsonl.gz 167 条 trial）。 data/04reward1successfulbuthackingfailed/ 也单独排除了 SWE-bench Pro Verified is exported under a separate benchmark path: The Pro collector continuously rerolls unresolved tasks. The current export has 138 verifier-correct…