Dharma Field Extractor's 2-billion-parameter model extracts selected document fields into JSON
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A small specialized model can perform a narrow, high-volume document-processing task more accurately and cheaply than general models on an organization's own hardware.
The model accepts a document image and either identifies structured fields itself or fills an exact schema with types, nesting, and tables. It is built on DharmaOCR-LITE v1.1 and trained with a LoRA adapter on 43,384 examples in English, Portuguese, and other languages. In Dharma's own test covering 546 documents and 1,638 scenarios, the model scored 0.834, above 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna. According to the developers' calculations, self-hosting on one L40S was 6–147 times cheaper than the compared APIs, depending on the model; this is an estimate of GPU-time cost, not a sales price. The model and a demonstration are available through Hugging Face.
Sources 1