MedCPT, INT8 ONNX
Dynamically quantized ONNX exports of NIH/NLM's MedCPT, for CPU inference through onnxruntime. Nothing here is a new model: the weights are MedCPT's, converted and quantized.
| directory | source | role |
|---|---|---|
medcpt-query |
ncbi/MedCPT-Query-Encoder |
encodes a query |
medcpt-article |
ncbi/MedCPT-Article-Encoder |
encodes a passage |
medcpt-cross |
ncbi/MedCPT-Cross-Encoder |
reranks a query/passage pair |
MedCPT is public-domain work of the US National Library of Medicine, trained on 255M PubMed search-log pairs. Credit belongs to its authors.
Quantization changes the numbers. These are INT8, not the original weights, so scores differ from the float model. On a 50-claim pooled evaluation the INT8 export scored 39/50 useful@10 against 40/50 for float -- close enough to ship, not identical. Do not treat these as a drop-in reproduction of published MedCPT results.
Medical disclaimer, carried forward from the original model cards: output is not intended for direct diagnostic use or medical decision-making without review and oversight by a clinical professional.