🔍 Read the full analysis: Dharma Field Extractor Explained: Document Data, Organized As JSON 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 identifying fields or following a user-provided schema. The company reports better benchmark quality and lower estimated costs than four tested frontier model configurations, but complete scores 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, either identifying fields on its own or filling a schema supplied by a user. The original analysis says it outperformed four tested frontier model configurations on its benchmark at substantially lower estimated cost, although the available summary does not include full scores or independent validation.
The model offers two ways to process a document. In open extraction, it identifies information from labels and page layout and returns key-value pairs. In schema-directed extraction, a user specifies field names and types; the model is intended to return those keys, place null where a requested value is absent, and organize tables row by row. Hugging Face says one checkpoint supports multiple document types without separate fine-tuning for each type.
The company describes potential uses such as extracting supplier names, invoice numbers, due dates and totals, or sending requisition details to scheduling systems. It says the model builds on DharmaOCR-LITE v1.1, a 2-billion-parameter OCR model, and uses a LoRA adapter trained on 43,384 examples. Those examples included English documents from the public web, internally sourced Portuguese documents from Brazil, and smaller collections in other European and Asian languages. The listed document categories include receipts, invoices, forms, bank statements, IDs and synthetic documents.
Hugging Face reports a benchmark of 546 documents and 1,638 cases, covering open extraction, partial schemas and schemas that included fields not present in the document. Its scoring combined field-level F1 by path, including table position, value similarity and the share of outputs that were valid JSON. The company says the quality gap over the four tested frontier configurations ranged from 1.8 to 9.3 points, but the supplied material does not specify the score scale or provide individual results.
The company also estimates costs at roughly six to 147 times lower than those configurations. Those estimates use average input and output tokens per page, including reasoning tokens, and providers’ published list prices. The available information does not give the extractor’s own price or enough detail to independently check the comparison.
From Scanned Pages to Business Records
Document processing often requires more than recognizing printed words. OCR transcribes page content, but a separate step is usually needed to assign a value to a field such as an invoice total or due date and format it for another system. A model that performs that mapping into schema-shaped JSON could reduce custom conversion work if its results are accurate on an organization’s documents.
The two request modes serve different needs: open extraction can help when the relevant fields are not known in advance, while a supplied schema fits software that expects fixed keys. Hugging Face’s reported cost advantage could also matter in workflows processing many pages. But estimates based on token use and list prices do not establish the total cost of deployment, which can include checks, corrections, integration and human review.
Products Worth Considering
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
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
A New Layer Beyond OCR
Hugging Face describes Field Extractor as an extension of DharmaOCR-LITE, which the company says it released in April for reading varied documents, including contracts, court filings, invoices, IDs and forms. The supplied material does not identify the year of that release or give a publication date for the new announcement.
The distinction is functional: OCR returns recognized text, while field extraction interprets that text and layout as named values, and can arrange table entries into rows. Hugging Face says it trained a LoRA adapter on field-level annotations across document types and languages. That is the company’s description of its development process; the material does not provide an independent evaluation of the training data or resulting performance.
“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 and Workflow Gaps
The benchmark summary leaves important details unresolved. It does not include full model scores, the score scale or individual results, and the reported comparisons have not been independently reproduced in the supplied material. Hugging Face’s cost estimates rely on measured average token use and providers’ published list prices; the extractor’s own pricing and the assumptions behind each configuration are not fully stated.
It is also unclear how performance varies across languages, document types and poor-quality scans, or how often users would need to correct errors. The benchmark does not establish results in any particular company’s workflow. Deployment requirements, availability of a public endpoint, handling of sensitive documents and the appropriate level of human review are not specified.
Evidence Needed Before Deployment
Hugging Face says users can try the model, but the supplied information does not give a release date, deployment instructions or a public endpoint. Further evidence that would help organizations evaluate it includes complete benchmark tables, reproducible testing details and results on document sets that differ from the evaluation data.
For prospective users, the practical next step is testing the model on their own document mix and checking how it handles missing values, tables and extraction errors. Decisions will also depend on pricing, privacy controls and review costs. Until those details and broader testing are available, the company’s benchmark and cost figures should be treated as reported results, not proof of reliability or savings in every setting.
Key Questions
What does Dharma Field Extractor do?
It processes document images and returns extracted information as JSON, either identifying fields itself or filling a schema provided by the user.
How is it different from OCR?
OCR recognizes and transcribes text. Field extraction adds a step by assigning values to named fields and, according to Hugging Face, organizing table data into rows.
How strong was its reported benchmark result?
Hugging Face reports a quality gap of 1.8 to 9.3 points over four tested frontier model configurations across 546 documents and 1,638 cases. The supplied summary does not state the score scale or provide individual scores, and the results have not been independently validated in the material provided.
Does the model guarantee accurate extraction or lower costs?
No such guarantee is established. The company reports lower estimated costs based on average token use and published list prices, but actual accuracy and total costs may depend on document quality, workflow integration, error correction and human review.
Primary source: Hugging Face · via ThorstenMeyerAI.com
