Baseline model trained on the T2KNOW corpus (Huntington's disease abstracts, 40 UMLS-derived labels, nested and same-span multi-label entities). Three training seeds are provided in seed_12345/, seed_23456/ and seed_34567/.
microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltextcode/ (Zenodo DOI above)model.pt is a PyTorch state_dict for the W2NER implementation in the release code (code/w2ner_baseline/). Build the model from w2ner_config.json and label_map.json (the label order must match exactly), then decode with nnw_threshold=0.7 and thw_threshold=0.7 (decoding.json), selected on the validation split. The default thresholds of 0.5 give different results.
Trained on the T2KNOW v1.0 document-disjoint train split (validation split for model selection). The v1.0 sentence-level split files contained duplicated sentence records; the corrected release v1.0.1 removes them without changing any annotation. No source text is included in this repository.
Overlap-aware scores from the official T2KNOW evaluator (partial span matches count half) on the de-duplicated T2KNOW v1.0.1 test split (151 documents, 1,413 sentences, 12,579 entities). Predictions are the archived outputs of these checkpoints.
| Seed | Precision | Recall | F1 |
|---|---|---|---|
| 12345 | 0.6032 | 0.9122 | 0.7262 |
| 23456 | 0.6002 | 0.9111 | 0.7237 |
| 34567 | 0.6022 | 0.9129 | 0.7257 |
Mean F1 over the three seeds: 0.7252 (population std 0.0011).
Scores measure agreement with the T2KNOW annotation layer (expert-checked, single annotator, see the paper), not semantic correctness. Very-low-support labels are unreliable.
MIT. The base model microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext is MIT-licensed (Hugging Face model metadata). T2KNOW annotations used for training are MIT-licensed; no third-party source text is included.
Repository: https://huggingface.co/edugredu/t2know-w2ner-biomedbert
seed_12345/model.pt 496 MBseed_23456/model.pt 496 MBseed_34567/model.pt 496 MBmicrosoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext