FAQ

My model only uses the accelerometer. Set channels: 3 in the encoder card; the gyroscope channels are dropped before your model. The row is marked accelerometer-only on the leaderboard.

My model expects 30 Hz input and 10 s windows. Set model_hz: 30; the harness resamples (anti-aliased when downsampling). Window lengths are fixed by the task cards (1 s to 10.24 s); a model that needs longer context sees the task window only.

My model normalises its own input. Set normalize_method: raw and do it inside forward. The benchmark fixes the raw input, not the normalisation; the released encoders use per-window z-scoring, physical-unit and raw inputs are supported as well.

Can I tune the probe or the fine-tuning recipe? No. The decoder, learning rate, budget and selection rule belong to the task card; imu-bench check compares every trained config with the card, and a cell that changed them is listed outside the ranking. Fine-tuning is the protocol for showing what the encoder reaches with adaptation.

Can I submit embeddings instead of an encoder? Not at the moment; see Submission modes.

I only ran some tasks. imu-bench package --allow-incomplete lays the results out and the pull request is accepted, but the row is not ranked until all twenty cells (ten tasks × two protocols) exist.

How long does a full run take? On one A100, a few hours for the twenty cells of a 21 M-parameter encoder; the kinematic fine-tunes dominate.

A reference row does not reproduce the released number. Check the data preparation first (Datasets); the reference rows control_random, control_raw, scratch and mae_5src exist so that a setup can be verified before a new model is trusted.

Which seed? 42 for every cell on the leaderboard. Differences inside the measurement noise are not separated; the paper reports the noise band of the kinematic tasks and bold marks ties at the printed precision.

Why is the leaderboard regenerated instead of edited? So that every number traces to a result file with its checkpoint step and config. imu-bench leaderboard rebuilds the tables from bench/results/.