Skip to content
Neural Daily

ModelsNews

Dharma Field Extractor: 2B model turns documents into JSON

🇭🇺 Olvasd el magyarul →

Dharma-AI's 2-billion-parameter model turns a document image into structured JSON, optionally following a schema you define. On the company's own benchmark it scored 0.834, ahead of every frontier model it tested.

By: Mikola Máté Patrik · · 3 min read

Dharma Field Extractor: a receipt image converted into structured JSON fields, Dharma-AI
Image: Dharma-AI · Press kit · Official cover image of the Dharma Field Extractor Hugging Face blog post · Source

What was released?

Dharma-AI introduced the Field Extractor (Hugging Face). It is a 2-billion-parameter model that takes a document image and returns structured JSON: exactly the fields you ask for from invoices, receipts, forms, IDs or contracts.

There are two modes. Without a schema, the model finds labels, values and tables on its own and returns everything as key-value pairs. With a schema, you set the key names and types (string, number, boolean), and the output follows it exactly, with no extra fields. If a requested field is not on the document, it returns null instead of inventing a value.

How was it trained?

The base is DharmaOCR-LITE v1.1, the company's 2B OCR-specialized model trained with SFT and DPO. On top of it sits a LoRA adapter trained on 43,384 examples: English multi-domain documents from the public web, internal Brazilian documents in Portuguese, and a smaller share of other European and Asian languages. One checkpoint handles every document type, with no per-type fine-tuning.

What does the benchmark show?

The company's own benchmark covers 546 documents and 1,638 cases across three scenarios: all fields with no schema, a partial schema, and a schema that includes deliberately nonexistent fields. The score is field-level F1, multiplied by Levenshtein similarity on matched values and by the share of valid JSON outputs.

The Field Extractor scored 0.834. Claude Opus 5 came closest at 0.816, followed by GPT-5.6 Sol at 0.811, Gemini 3.7 Flash at 0.792 and GPT-5.6 Luna at 0.741. The frontier models ran with reasoning set to low.

Dharma-AI chart: extraction quality, Field Extractor 0.834 vs Claude Opus 5, GPT-5.6 Sol, Gemini 3.7 Flash, GPT-5.6 Luna
Dharma-AI · Press kit · Official benchmark chart from the Dharma Field Extractor blog post · Source

Source: Dharma-AI. Extraction quality on the company's 546-document benchmark.

The cost gap is wider. According to Dharma-AI, 1,000 extractions cost $0.26 on a single self-hosted L40S GPU priced at $2.24 per hour. Frontier APIs at list price range from $1.55 (GPT-5.6 Luna) to $38.25 (GPT-5.6 Sol). The company notes this is not a sales price, since self-hosted cost depends on hardware, volume and latency needs.

Dharma-AI chart: cost per 1,000 extractions, Field Extractor $0.26 vs frontier APIs from $1.55 to $38.25
Dharma-AI · Press kit · Official cost chart from the Dharma Field Extractor blog post · Source

Source: Dharma-AI. Cost per 1,000 extractions on a logarithmic scale.

What it means for developers

If you extract fields from invoices, forms or IDs at high volume, a small task-trained model can be cheaper than prompting a frontier API. The numbers come from the company's own benchmark, though, and the Field Extractor weights are not public: the model can be tried on the Dharma-AI portal after signing in. The underlying DharmaOCR-LITE is available on Hugging Face.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles