Result files

One JSON per cell at bench/results/<task>/<encoder>__<protocol>__s<seed>.json, written by the task's scorer:

{
  "task": "microgesture_openwatch",
  "encoder": "mae_5src",
  "protocol": "probe",
  "seed": 42,
  "metrics": {"clip_macro_f1": 0.647, "clip_acc": 0.678, "window_macro_f1": 0.494, "window_acc": 0.536},
  "primary_metric": "clip_macro_f1",
  "eval_split": "test",
  "run": "bench_microgesture_openwatch__mae_5src__probe__s42",
  "checkpoint": "runs/.../checkpoints/best.pt",
  "checkpoint_step": 1200,
  "selection": {"split": "val", "monitor": "macro_f1", "mode": "max"},
  "job_ids": [7212525],
  "provenance": {"legacy_run": false, "scorer": "bench.score.microgesture", "scored_at": "..."},
  "status": "ok",
  "notes": "",
  "schema_version": 1
}

Field	Meaning
metrics	every metric the scorer computes; the leaderboard reads primary_metric
eval_split	test where a test split exists (gesture, retrieval), otherwise the held-out validation split the card defines
checkpoint_step	the step selected by the card's rule on the validation split
provenance	scorer, time, and whether the run predates the harness (legacy_run)

imu-bench leaderboard collects these files into bench/leaderboard_tracks.json (per track and protocol: rows, cells, ranks) and the Markdown/HTML tables. The released result files, the registry and the leaderboard data are mirrored in the dataset YiquanLi/IMU-Bench-results, whose results_long.csv lists every cell as one row.