TermMap-X+ (xplus-sf): trained weights
Weights for the TermMap-X+ subconcept-free system (clinical mention + chart chunk -> one ICD-10-CM code and one
SNOMED CT concept). Code and instructions: see the GitHub repository's README. Pull with python setup_model.py.
| path | what | base model |
|---|---|---|
models/sapbert-ft/ |
SapBERT fine-tuned (InfoNCE) on training mentions -> dictionary names | cambridgeltl/SapBERT-from-PubMedBERT-fulltext |
models/cross-ctx/ |
X's cross-encoder, mention + chart context, listwise 1+7 | microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext |
models/cross-noctx/ |
X's cross-encoder, mention only | microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext |
data/processed/dense_*.npy |
SapBERT-ft embeddings of every dictionary alias (fp16) | - |
Not included: the char TF-IDF + BM25 alias indexes (data/processed/lexical_*.pkl); termmap rebuilds them from the
alias tables on first run. Also not included (public, pulled at pinned revisions by setup_model.py): Qwen/Qwen2.5-1.5B-Instruct (SNOMED gate),
Qwen/Qwen3-Reranker-0.6B (ICD zero-shot scorer). Every file is listed with its sha256 in WEIGHTS_MANIFEST.json.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support