[professor3333]

Summary

reports/neural/run-3/: the third GPU run (Kaggle background version, one T4).

• Grid: 24 runs, up to 20 epochs, patience 3, min_epochs 6.
• Selection: early stopping and each grid's best run used only the selection half of the model-drafted dev records (splits/selection-v1/selection.json, 111 docs).
• Comparisons: on the held-out half (61 docs, 46 other groups) and on the 10 hand-labelled documents.

Results:

• DeBERTa-v3-small (sqrt, 5e-5): the seed mean is 0.197 ± 0.011 held-out typed F1, up from 0.131 in run 2.
  • The compared run reaches 0.185 [0.145, 0.235], against 0.127 [0.054, 0.201] for crf-batches-01-02: paired +0.058 [−0.008, +0.134], not clear.
  • Its false-prerequisite rate is +0.072 [−0.070, +0.196] against the CRF, not clear.
• Uniform weights at 5e-5 now train (DeBERTa 0.145 ± 0.014) and give the lowest transformer false-prerequisite rates. At 2e-5 they still predict nothing by epoch 6.
• DistilBERT: best 0.099; span precision clearly worse than the CRF's.
• Hand-labelled slice: no clear differences.
• Committed: per-run metadata, all reports (plus the CRF's held-out report), and the predictions of the two compared runs. README results now show the held-out comparison instead of run 2's selection-biased rows; the model card is updated.

The test split is not touched.

Checks

• ruff check, ruff format --check, pytest -m "not slow" (241 passed twice); the reproduce command matches the report.
• One local run aborted in native code (libc++abi: recursive_mutex lock failed) before reporting. Two re-runs were clean. It is intermittent and possibly linked to memory pressure on the development Mac; this is not confirmed.