Code-only release — no data changes from v3.1.1.
Highlights (v3.2.0 + v3.2.1)
- Unified PubMedBERT model for all sections (AR, BW, WP) replaces three separate models, with per-section classification thresholds
- TAC F1: 89.87 (up from 87.54), Recall: 94.68 (up from 84.08)
- Web-based annotation tool for creating independent reference sets (#31)
- 30 independently annotated FDA labels as new hold-out evaluation set
- Snakemake training workflow with 44 experiments across 10 BERT variants (#81)
- Checkpoint/resume support for interrupted training runs (#30)
- Curation flag experiments: filtering does not improve performance (#26)
- Reference set characterization with rebuilt Denmer-Fushman annotations (#27)
v3.2.1 patch
- Fix FK violation in
vocab_rxnorm_ingredient_to_product— column order mismatch in Postgres/MySQL schema caused foreign key violations when loading CSVs (#82)
Data
This release does not include a data update. The most recent data release is v3.1.1.