Dhisanth/hilabs-term-mapping

HiLabs clinical term mapping engine — trained models

Trained artifacts for the HiLabs AI Quest 2026 term-mapping engine (clinical entity + chart context → ICD-10-CM and SNOMED CT). Code, instructions and write-up: the GitHub repository linked in the submission.

The engine downloads this repository automatically into models/ on the first python run.py predict.

Path	Contents
biencoder/, biencoder_r1/	SapBERT bi-encoders fine-tuned contrastively (round 2 / round 1 hard negatives)
crossencoder/	SapBERT cross-encoder reranker (term + chunk context vs candidate)
llm_adapter/	QLoRA adapter for unsloth/Qwen3-4B-Instruct-2507-bnb-4bit (listwise reranker)
r1_*, r2_*, r2l_*, r3_* __bag*.txt	5-seed bagged LightGBM LambdaRank rankers; ranker_features.json lists their features
ctx_*.pkl	chunk-context concept-type classifiers (vocabulary stored as hashes, no plain chart text)
pairs.csv	ICD↔SNOMED code co-occurrence pairs (codes only)

No training data, chart text, dictionaries or API keys are included. Base models: SapBERT (cambridgeltl/SapBERT-from-PubMedBERT-fulltext, Apache-2.0) and Qwen3-4B-Instruct-2507 (Apache-2.0).