Evaluation logs for the Bench2Dex leaderboard entry `pyrua-dex` (PKU DAGroup), a coding-agent system that uses no human demonstration data. GitHub PR: https://github.com/Bench2Dex/Bench2Dex-Leaderboard/pull/7
- `pyrua-dex/README.md`, `pyrua-dex/leaderboard_pyrua-dex.json` (identical to the GitHub results JSON)
- `pyrua-dex/per_task_eval/<task>/<None|Equi|Inv|Full>/per_episode.jsonl` and `summary.json`: the original Bench2Dex outputs of each channel's 50 episodes; `native_shards/000` and `001` hold the two batches of 25 as Bench2Dex wrote them.
- Sanitization: absolute asset paths are made relative (`dex2bench_dataset/...`); nothing else is changed (structure, fields, units, seeds, episode counts as written).
- Coverage: 26 tasks x 4 channels x 50 episodes = 5,200 episodes; Bench2Dex `ae870fb0e1f6237f20c8302198ee3850118892e1`. Overall SR 44.62, LSCR 67.02, Eff 0.8889 (Score 59.126).
π€ Generated with [Claude Code](https://claude.com/claude-code)