CLI

imu-bench is installed by pip install -e . (entry point bench.cli:main).

Command	Does
imu-bench tasks	lists the ten task cards: dataset, split, input, metric
imu-bench eval --encoder <id or card.yaml> [--tasks …] [--protocols probe finetune] [--seed 42] [--slurm] [--force] [--dry-run]	registers a card file under bench/registry/encoders.d/, then trains and scores every requested cell, here or as SLURM chains
imu-bench check <id>	completeness (20 cells), protocol fidelity (trained configs identical to the task cards), card completeness, track eligibility
imu-bench package <id> [--out submission] [--allow-incomplete]	lays out results/, registry/<id>.yaml, configs/ and check.json for the leaderboard pull request
imu-bench leaderboard [--site index.html] [--readme README.md] [--quiet]	regenerates bench/leaderboard_tracks.json and the two track tables; optionally rewrites the website and README blocks
imu-bench audit <dir> --subject-regex '<pattern>' [--glob '**/*.csv'] [--acc-cols AccX AccY AccZ] [--threshold 0.999]	finds duplicated, mirrored or rotated participants in a dataset release
imu-bench fetch-weights [ids…]	downloads the released weights of the open-track models to where their cards expect them

The lower-level modules remain available: python -m bench.make_config, python -m bench.launch, python -m bench.score, python -m bench.leaderboard (see Reproducing the paper).

Environment variables:

Variable	Meaning
IMUBENCH_DATA_ROOT	directory holding the datasets in the layout of Datasets; every dataset path in the cards is rewritten to it
IMUBENCH_VENV	environment directory activated by generated SLURM scripts (default imu-bench/)
BENCH_PARTITION, BENCH_ACCOUNT	SLURM partition and account for --slurm