Files
Khalil Meftah f1efa588b8 fix(sarm): fail fast on unusable episode annotations (#4306)
* fix(sarm): warn when dense/sparse targets silently collapse to all-zero

In dense_only/dual modes, if meta/episodes/*.parquet has no usable
subtask columns (column absent or NaN), _load_episode_annotations
returns None and find_stage_and_tau yields stage 0 / tau 0 for every
frame. Training "succeeds" but the head silently learns to predict 0
everywhere, with no warning. This complements #2880 (which restored
loading of episodes_df): there the DataFrame is loaded but the
*_subtask_names column is missing/NaN.

Add a one-time validation at processor construction that logs a clear
warning (all episodes missing -> predict-all-zero; some missing ->
partial). Purely additive logging, no change to training math.

Closes #3842

* fix(sarm): fail fast on unusable episode annotations

---------

Co-authored-by: 1thanShih <Smartshithan1620.en12@nycu.edu.tw>
2026-08-03 13:30:00 +02:00
..
2026-04-28 17:56:24 +02:00