Add Phoenix-MWA 20k evaluation logs and metric revision
Add Phoenix-MWA (w. Pretrain) evaluation logs
Policy ID: phoenix-mwa.
GitHub leaderboard PR: https://github.com/Bench2Dex/Bench2Dex-Leaderboard/pull/5
This package includes 26 tasks × 4 channels × 50 episodes, matching README/results JSON, sanitized original native logs under per_task_eval/, and the Task09 revision under revisions/task09_axis_y/. Source paths in the results JSON explicitly identify the selected score inputs. Additional revision copies are not counted as additional episodes.
Reported scores: SR 45.5961538%, LSCR 65.6080128%, efficiency 0.9025351 (25 tasks with success). Checkpoint mapping, episode seeds, execution budgets, historical Task27/42 differences and Task09 revision details are recorded in the evaluation metadata. Numerical Task09 replay inputs and a portable verification script are included.
Target branch: main. The accepted HF commit can be pinned in evaluation.logs.revision after review.