Dharma Field Extractor: Any Document In, Structured Fields in JSON Out
Dharma-AI has announced Dharma Field Extractor, a 2-billion-parameter model that turns a document image into structured JSON. It offers an open mode and a schema-directed mode.
What the article says
The open mode finds key-value pairs without a predefined list. The schema-directed mode lets users set the keys and types (string, number or boolean); missing fields return null instead of guessed values. The model is a LoRA adapter built on DharmaOCR-LITE v1.1, trained on 43,384 examples in English, Brazilian Portuguese and other languages.
On a 1,638-case benchmark it scores 0.834, above Claude Opus 5 (0.816) and GPT-5.6 Luna (0.741). The authors state a self-hosted cost 6 to 147 times lower than the frontier APIs compared, based on GPU time on a single L40S at $2.24 per hour. The tool can be tried through a web portal; the article does not say whether the weights are publicly released.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Hugging Face. The content and rights belong to the publisher: please read the original article.