THIRAWAT-BioLORD

THIRAWAT-BioLORD is a fine-tuned ColBERTv1-style late-interaction reranker for drug terminology mapping to standardized OMOP Drug concepts, especially RxNorm and RxNorm Extension.

THIRAWAT stands for Terminology Harmonization using Late-Interaction Reranker With Alignment-tuned Transformers. This model is designed to rerank candidate drug concepts generated by a first-stage retriever, not to perform full terminology mapping by itself.

This checkpoint is initialized from BioLORD-2023, a biomedical sentence embedding model trained to produce semantic representations for clinical sentences and biomedical concepts using definitions, biomedical knowledge graph information, and LLM-generated clinical terminology definitions.

It is best used through THIRAWAT Mapper:

https://github.com/sidataplus/THIRAWAT-mapper

Evaluation Summary

THIRAWAT-BioLORD was evaluated as part of the full THIRAWAT Mapper pipeline. In plain terms:

• Hits@1 means the correct mapping was ranked first.
• Hits@3 means the correct mapping appeared in the top 3 suggestions, which is useful for human review.

Across the evaluated datasets, the full BioLORD-based pipeline ranked the correct drug concept first in 82.0%–85.8% of cases, and placed it within the top 3 in 87.1%–93.4% of cases.

The table below compares the full SapBERT-based and BioLORD-based THIRAWAT Mapper pipelines using reranking plus deterministic tie-breaking.

Dataset	Pipeline / Reranker	Correct at #1	Correct in top 3	Difference vs THIRAWAT-SapBERT at #1	Difference vs THIRAWAT-SapBERT top 3
Branded Drugs	THIRAWAT-SapBERT	94.2%	96.4%	baseline	baseline
Branded Drugs	THIRAWAT-BioLORD	82.0%	87.1%	-12.2 pp	-9.3 pp
Clinical Drugs	THIRAWAT-SapBERT	85.9%	93.0%	baseline	baseline
Clinical Drugs	THIRAWAT-BioLORD	85.2%	92.3%	-0.7 pp	-0.7 pp
Thai Medicines Terminology	THIRAWAT-SapBERT	86.8%	95.4%	baseline	baseline
Thai Medicines Terminology	THIRAWAT-BioLORD	85.8%	93.4%	-1.0 pp	-2.0 pp

These are full-pipeline results, not standalone checkpoint-only results. The BioLORD-based pipeline used BioLORD-2023 retrieval, THIRAWAT-BioLORD reranking with BiMaxSim, and deterministic tie-breaking. The deterministic tie-breaker is implemented in THIRAWAT Mapper, not inside this Hugging Face model checkpoint.

In this evaluation, THIRAWAT-SapBERT was stronger overall, especially for Branded Drugs. THIRAWAT-BioLORD remained competitive on Clinical Drugs and Thai Medicines Terminology, where the gap at Hits@1 was approximately 1 percentage point or less.

Paper

This model is described in:

Efficient Drug Terminology Mapping with Bidirectional Late-Interaction Reranking and Deterministic Reordering Healthcare Informatics Research, 2026;32(2):156–165 DOI: https://doi.org/10.4258/hir.2026.32.2.156

Please cite the paper if you use this model or THIRAWAT Mapper.

Intended Use

Use this model for reranking candidate OMOP/RxNorm drug concepts in medication terminology mapping workflows.

Typical pipeline:

1. Retrieve candidate concepts with BioLORD-2023 or another biomedical retriever.
2. Rerank candidates with THIRAWAT-BioLORD.
3. Score late-interaction matches with inference-time BiMaxSim.
4. Apply deterministic tie-breaking for near-ties using drug-specific cues such as strength, dosage form/route, release characteristics, and brand annotations.
5. Review the final mappings before production use.

This model is intended for terminology mapping and mapping assistance. It is not a prescribing tool, medication safety checker, or standalone clinical decision support model.

Example: THIRAWAT Mapper CLI

pip install thirawat-mapper

Interactive query:

thirawat infer query \
  --db data/lancedb/db \
  --table concepts_drug \
  --device auto \
  --reranker-id sidataplus/THIRAWAT-BioLORD

Bulk mapping:

thirawat infer bulk \
  --db data/lancedb/db \
  --table concepts_drug \
  --input data/usagi.csv \
  --out runs/mapping \
  --candidate-topk 200 \
  --n-limit 20 \
  --device auto \
  --reranker-id sidataplus/THIRAWAT-BioLORD

Model Details

• Base model: FremyCompany/BioLORD-2023
• Architecture: ColBERTv1-style late-interaction reranker
• Model type: PyLate ColBERT model
• Training objective: one-sided MaxSim
• Recommended inference scoring: BiMaxSim
• Projection dimension: 128
• Max query length: 96 tokens
• Max candidate length: 96 tokens
• Primary domain: OMOP Drug mapping to RxNorm / RxNorm Extension

Relationship to THIRAWAT-SapBERT

THIRAWAT-BioLORD and THIRAWAT-SapBERT use the same THIRAWAT late-interaction reranking approach, but start from different biomedical encoders.

• THIRAWAT-SapBERT is initialized from SapBERT-XLMR, which is optimized for biomedical entity linking and synonym alignment.
• THIRAWAT-BioLORD is initialized from BioLORD-2023, which emphasizes biomedical semantic similarity grounded in definitions and knowledge graph descriptions.

In the reported drug-mapping evaluation, THIRAWAT-SapBERT achieved the strongest overall full-pipeline results. THIRAWAT-BioLORD is provided as an alternative checkpoint with a different biomedical representation bias and may be useful for comparison, ablation, or deployment settings where BioLORD-style semantic representations are preferred.

Limitations

• Evaluated primarily for drug terminology mapping.
• Not yet validated for other OMOP domains such as Condition, Procedure, Measurement, or Observation.
• Performance depends on the target vocabulary and ATHENA/RxNorm coverage.
• BioLORD-based retrieval showed lower retriever-only performance than SapBERT-XLMR in the reported drug-mapping setting, although reranking and deterministic tie-breaking substantially narrowed the gap.
• Local brands may be mapped to generic ingredient-strength-form concepts when no corresponding branded target exists.
• Remaining errors often involve clinically close candidates, such as strength, form, release, combination-product, or brand-coverage mismatches.
• Automated mappings should be reviewed before use in production OMOP ETL or research workflows.

Citation

@article{adulyanukosol2026thirawat,
  title = {Efficient Drug Terminology Mapping with Bidirectional Late-Interaction Reranking and Deterministic Reordering},
  author = {Adulyanukosol, Natthawut and Chaisutyakorn, Krittaphas and Sombutjaroan, Saknarong and Kanjanapong, Suchanan and Suriyaphol, Prapat},
  journal = {Healthcare Informatics Research},
  year = {2026},
  volume = {32},
  number = {2},
  pages = {156--165},
  doi = {10.4258/hir.2026.32.2.156}
}

If you use the BioLORD-2023 base model directly, please also cite the original BioLORD-2023 paper.

@article{remy2024biolord,
  title = {BioLORD-2023: semantic textual representations fusing large language models and clinical knowledge graph insights},
  author = {Remy, François and Demuynck, Kris and Demeester, Thomas},
  journal = {Journal of the American Medical Informatics Association},
  year = {2024},
  volume = {31},
  number = {9},
  pages = {1844--1855},
  doi = {10.1093/jamia/ocae029}
}