🔍 Read the full analysis: Dharma Field Extractor: Any Document In, Structured Fields In JSON Out on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Hugging Face has introduced Dharma Field Extractor, a 2-billion-parameter model that converts document images into structured JSON, either by finding fields itself or following a user-provided schema. The company reports higher benchmark quality and lower estimated costs than four tested frontier-model configurations, but full scores, pricing details and independent validation are not provided.
Hugging Face has introduced Dharma Field Extractor, a 2-billion-parameter model that turns document images into structured JSON for tasks such as invoice processing, as described in the original analysis. The company says the model performed better on its benchmark than four tested frontier-model configurations and cost less to run, though the supplied information does not include full scores, the model’s own pricing or independent test results.
The extractor offers two ways to request information. In open extraction, it identifies fields from the document’s labels and layout and returns key-value pairs. In schema-directed extraction, a user supplies field names and types, and the model is intended to return those keys, set absent values to null and organize table entries row by row.
Hugging Face says a single checkpoint handles different document types without per-type fine-tuning. Its examples include extracting suppliers, invoice numbers, due dates and totals, as well as passing requisition details to scheduling systems. These are described as potential applications; the source does not report deployment results from customers or a particular organization.
The model is built on DharmaOCR-LITE v1.1, a 2-billion-parameter OCR model, with a LoRA adapter trained on 43,384 examples, according to Hugging Face. The examples covered English documents from the public web, internally sourced Portuguese documents from Brazil, and smaller collections in other European and Asian languages. Document types included receipts, invoices, forms, bank statements, identification documents and synthetic samples.
From Scanned Pages to Business Records
Document extraction addresses a step beyond reading text: converting page content into named fields that other software can use. OCR can transcribe words, but a business system may still need to identify which number is the invoice total, which date is the due date, or how values in a table map to columns. A tool that reliably returns the requested structure could reduce custom mapping and manual data entry.
The two request modes target different needs. A fixed schema may suit a system that expects particular keys, while open extraction could help when fields vary or are not known in advance. Hugging Face’s reported cost advantage may attract organizations handling large page volumes, but the figures are estimates, not evidence of total savings in a live workflow. Actual costs can also include checking errors, processing difficult documents and reviewing sensitive records.
The announcement matters as a specific automation proposal, not proof that organizations can remove human review. The source does not establish performance across individual businesses, document collections or error-handling processes.
Products Worth Considering
EcoFlow Delta 3 Portable Power Station – 1024 Wh, 1800 W
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
Anker SOLIX C300 Portable Power Station, 288Wh
As an affiliate, we earn on qualifying purchases.
Built on Hugging Face OCR Work
Hugging Face describes the Field Extractor as an extension of its OCR work. The company says it released DharmaOCR-LITE in April for reading varied documents, including contracts, court filings, invoices, identification documents and forms. The source does not specify the year of that release or give a publication date for this announcement.
The distinction is functional: OCR follows the text on a page, while field extraction assigns values to requested labels and may match table entries to columns. Hugging Face says its adapter was trained on field-level annotations across document types and languages. That background helps explain the model’s intended role, but training coverage alone does not establish accuracy on unseen documents or languages.
“The Field Extractor takes a document image and either a request for everything it can find, or an exact schema to fill, and returns structured JSON.”
— Hugging Face
Benchmark Results Need More Detail
Hugging Face says its benchmark included 546 documents and 1,638 cases, covering open extraction, partial schemas and schemas with fields that did not appear in the documents. The company says quality differences against four frontier-model configurations ranged from 1.8 to 9.3 points. The supplied material does not identify the score scale, provide individual results or include the full benchmark tables, so readers cannot assess the comparison in detail.
The company also reports costs roughly six to 147 times lower than those tested configurations. Its estimates use average input and output tokens per page, including reasoning tokens, and providers’ published list prices. The source does not state the extractor’s own price or fully specify the assumptions needed to reproduce the comparison. It says the other models used low reasoning settings, with adaptive effort for Claude Opus 5.
Independent reproduction and results on varied, unseen documents remain unavailable in the supplied information. Also unclear are error rates by language and document type, performance on poor-quality scans, the amount of human review needed, and how the model handles sensitive documents in deployment. The benchmark does not establish that the model will perform similarly in a particular organization’s workflow.
Deployment and Independent Testing
Hugging Face says users can try the model, but the supplied material does not provide a release date, deployment requirements or details of a public endpoint. Further evidence would help potential users judge whether the reported results translate into practical use, including complete benchmark tables, reproducible evaluation details and tests by organizations with different document formats and field requirements.
For teams considering adoption, the next steps would include testing on their own documents, deciding how to flag uncertain or missing values, and measuring the cost of human review alongside model use. Until those details and independent results are available, the announcement supports a possible route from document images to structured records, but does not settle how reliably or economically it will work across business settings.
Key Questions
What does Dharma Field Extractor do?
It takes a document image and is designed to return information as structured JSON. Users can ask it to identify available fields or supply a schema of requested fields.
How does schema-directed extraction work?
A user specifies field names and types. Hugging Face says the model is intended to return those keys, use null when a requested value is absent, and arrange table data row by row.
What did Hugging Face report in its benchmark?
The company says the evaluation covered 546 documents and 1,638 cases, and that the model’s quality was 1.8 to 9.3 points higher than four tested frontier-model configurations. Full scores and the scoring scale are not included in the supplied material.
Are the cost savings independently confirmed?
No independent confirmation is included. Hugging Face estimates costs were roughly six to 147 times lower than for the tested configurations, using average token use and published list prices. The extractor’s own price and the full calculation details are not provided.
Can businesses use it without human review?
The supplied information does not establish that. Error rates, performance on organizations’ own documents and the amount of human review required remain unclear, so teams would need to test the model in their intended workflow.
Primary source: Hugging Face · via ThorstenMeyerAI.com
