| --- |
| license: apache-2.0 |
| language: en |
| base_model: dmis-lab/biobert-base-cased-v1.1 |
| tags: |
| - biomedical |
| - named-entity-recognition |
| - nested-ner |
| - huntingtons-disease |
| --- |
| |
| # t2know-biobert: BioBERT span classifier (flat, multi-label BIO) |
|
|
| 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/`. |
|
|
| - Base model: [`dmis-lab/biobert-base-cased-v1.1`](https://huggingface.co/dmis-lab/biobert-base-cased-v1.1) |
| - Corpus: T2KNOW v1.0.1 (Zenodo DOI: [10.5281/zenodo.23801223](https://doi.org/10.5281/zenodo.23801223)) |
| - Paper: under review; the citation will be added on publication. |
| - Code: release archive `code/` (Zenodo DOI above) |
|
|
| ## Usage |
|
|
| The checkpoint uses the custom `BertForSpanClassification` head from the release code (`code/flat_baselines/modeling.py`); it cannot be loaded with a stock `AutoModel` head. Decode with `logit_threshold=0.5` and `start_confidence=0.9` (`decoding.json`), which were selected on the validation split. |
|
|
| ## Training data |
|
|
| 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. |
|
|
| ## Test results |
|
|
| 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 | 0.5233 | 0.7396 | 0.6130 | |
| | 23456 | 0.5309 | 0.7383 | 0.6177 | |
| | 34567 | 0.5323 | 0.7289 | 0.6153 | |
|
|
| Mean F1 over the three seeds: 0.6153 (population std 0.0019). |
|
|
| ## Limitations |
|
|
| Scores measure agreement with the T2KNOW annotation layer (expert-checked, single annotator, see the paper), not semantic correctness. Very-low-support labels are unreliable. |
|
|
| ## Licence |
|
|
| Apache License 2.0, following the base model: BioBERT is released under Apache 2.0 (https://github.com/dmis-lab/biobert, LICENSE; the Hugging Face model page carries no licence tag). T2KNOW annotations used for training are MIT-licensed; no third-party source text is included. |
|
|
| Repository: `https://huggingface.co/edugredu/t2know-biobert` |
|
|
