Whisper Small learned, but missed the quality target.

The corrected sweep

All three ranking arms ran for 1,500 steps. The selected rate, 1e-3, was then trained fresh for 4,000 steps. Smaller rates performed worse within the ranking budget. This single-seed sweep does not establish a universally optimal rate.

The final’s WER was 0.0758 percentage points higher than the short reference. These fresh runs had different linear-decay budgets, so this does not isolate the effect of 2,500 additional updates.

At step 1,200, evaluation WER rose to 99.17% and recovered to 33.72% at step 1,400. The cause was not isolated. Late checkpoints fluctuated around 31–33% WER.

Performance was uneven.

The worst-scoring evaluation speaker had 62.70% WER across 37 recordings. A 25-row text review found phrase substitutions and spelling disagreements, but listening and the broader A13 data audit remain pending. Text coincidences alone do not prove incorrect audio alignment.

What we carry into the next run

1. L1 · Check the real data path. Keep label-prefix assertions, save/reload checks, and pinned scoring.
2. L3 · Judge recognition, not training loss alone. Low training loss did not translate into lower late-stage WER.
3. L4 · Capture anomalies deliberately. Batch identities, gradient norms, and loss-scaler changes are proposed checks for a relevant repeat, not completed instrumentation.
4. L5 · Audit before choosing the intervention. A13 will review at least 100 errors and roughly 500 training examples.

Download and provenance

Audited locally against the retained private run artifacts. This public snapshot contains aggregate results and checkpoint curves, not audio or per-example transcripts.

No dataset audio, per-example transcripts, private host details, or credentials are included in this public results snapshot.