Sorts a document image into 27 classes and flags medical and handwritten, in ~96M params.
An encoder-only model on a SigLIP2-base backbone. One pass sorts an image into one of 27 classes across two levels - lab report, discharge summary, ABHA card, Aadhaar, invoice - and says whether it is medical and whether it is handwritten. The int4 build is 60 MB and runs on CPU.
Benchmarks
| Metric | Value | Dataset |
|---|---|---|
| Accuracy, 27 classes (%) (higher is better) | 96.45 | Eka internal test set: 16,204 images |
| Macro-F1 (higher is better) | 0.965 | Eka internal test set: 16,204 images |
Install
Download from Hugging Face
hf download ekacare/med-doc-classifier Links
Topics
document-classificationimage-classificationsiglip2medical-recordscpu