Add MXI: full 26-task evaluation logs
MXI — independently executed evaluations of our fine-tuned policies
This update provides the standalone results requested by the reviewer. All 26 task policies were fine-tuned by 1011.ai. No unchanged released Bench2Dex task checkpoint is included. It replaces the earlier mixed-origin submission.
The evaluations were independently executed by 1011.ai: 26 tasks, four channels, 50 episodes per channel, 5,200 complete episodes. Each task uses its respective fine-tuned checkpoint, fixed across all four channels. These are our completed native evaluation runs; independent execution does not mean a fresh held-out confirmation experiment. The metadata records checkpoint identities and the scope of this result.
| Metric | Value |
|---|---|
| SR | 36.865385% |
| LSCR | 54.783654% |
| Efficiency | 0.970366548 |
| Score / MR | 50.945733 |
The corresponding complete native logs and PNG logo are included. Submission-validator checks pass; maintainer acceptance is pending.
Contributor: 1011.ai · Contact: lizhehan@1011ai.cn
GitHub metadata PR: https://github.com/Bench2Dex/Bench2Dex-Leaderboard/pull/6
Hugging Face logs PR: https://huggingface.co/datasets/Bench2Dex/Leaderboard/discussions/6