Veriwire
← Back to the wire

Dharma Field Extractor: Any Document In, Structured Fields in JSON Out

AnnouncementModelNov 4, 2026

Dharma announced Dharma Field Extractor, a 2-billion-parameter model built on DharmaOCR-LITE v1.1 that returns structured JSON fields from any document image, either open-ended or schema-directed, with null for missing fields. Trained via a LoRA adapter on 43,384 field-annotated examples across English, Brazilian, and multilingual data. On a 546-document, 1,638-case benchmark, it outperformed four frontier LLM configurations by 1.8 to 9.3 F1 points at roughly 6 to 147 times lower cost.

Why it matters

Dharma has released Dharma Field Extractor, a compact language model built on its DharmaOCR-LITE checkpoint and specialized for pulling specific structured fields out of document images. Unlike general OCR, which returns all text on a page, the model returns typed JSON: users can request every field found on a document or supply an exact schema, and the model fills it precisely, returning null for fields that are absent rather than inventing values. The same checkpoint handles invoices, medical requisitions, IDs, contracts, and forms across English, Brazilian, and other language data, with no per-document-type tuning. Dharma benchmarked it against four frontier LLM configurations across open, partial-schema, and deliberately misleading-schema scenarios, scoring field-level F1 weighted by value similarity and JSON validity. The extractor beat all four frontier models on quality while running at substantially lower cost per page. The company frames this as a targeted, cheaper alternative to prompting general-purpose APIs for high-volume field extraction.

Receipt № 27271 source · awaiting confirmation ◐

Evidence

1source· awaiting independent confirmation

No score is assigned. Sources and their independence are shown in the citation chain below.

Citation chain · 1 source

01med
huggingface.coBloghttps://huggingface.co/blog/Dharma-AI/dharma-field-extractorThis is Dharma's own announcement, detailing the model's two request modes, its field-annotated multilingual training data, and its benchmark methodology against frontier LLMs.
PRIMARY
DharmaCompanyDharma Field ExtractorModelDharmaOCR-LITEModel

Related

Nov 4Tencent WeVisOne-VL Unifies Detection, OCR, Segmentation and MattingNov 3Palette: teaching a small model to build decks in your design languageNov 3The 'Astra can read and write weights to improve Qwen' post is massively misleading · honnibal.dev
Canonical: https://huggingface.co/blog/Dharma-AI/dharma-field-extractor