[professor3333]

Summary

• Run 1 record (reports/neural/run-1/): the first GPU run, 12 runs on a Kaggle T4.
  • Every run over-predicts spans: 39–47 conditions per document (CRF 1.8), span precision ≈ 0.015.
  • Best typed F1 against the 172 model-drafted dev records is 0.023, against 0.145 for crf-batches-01-02. The paired difference is clear (−0.122 [−0.160, −0.081]).
  • Reports, per-run metadata (per-epoch losses and dev scores) and the archive SHA-256 are committed; the 24 MB of predictions are not.
• Diagnosis:
  • Balanced (inverse-frequency) span-tag weights make a missed B about 65× as costly as a wrong one.
  • With about 31 optimizer steps per epoch, most runs were still improving at epoch 6.
• Changes for run 2:
  • TrainConfig.bio_weighting takes inverse (run 1), sqrt (new default) or none.
  • The DistilBERT and DeBERTa grids try sqrt and none × 2 learning rates × 3 seeds (24 runs), with up to 15 epochs and patience 3.
  • The grids are renamed distilbert-2 / deberta-2, so resumed runs never reuse run-1 bundles.
  • run_neural.sh reads run names from the configs and predicts and packs only the current grids. score_neural.sh groups runs by grid and clears older predictions.
  • Docs and the model card are updated.
• Tests: weighting modes, the new grid sizes. The real run-1 archive was scored with the updated scoring script.

The test split is not touched.

Checks

• ruff check, ruff format --check, mypy src/, pytest -m "not slow" (238 passed); shellcheck clean.
• One earlier local full-suite run reported 1 failure that did not reproduce in four further runs, and the failing test was not captured. CI is the check here.