🔍 Read the full analysis: From Unstructured Documents To Structured Fields With Dharma 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 turns document images into structured JSON, either by finding fields itself or following a user-provided schema. The company reports higher benchmark quality and substantially lower estimated costs than four tested frontier model configurations, but has not supplied full scores, pricing details or independent validation.
Hugging Face has introduced Dharma Field Extractor, a 2-billion-parameter model that converts document images into structured JSON for use in software and business workflows. The company says the model outperformed four tested frontier model configurations on its benchmark at substantially lower estimated cost, though the supplied results lack full score tables and independent validation.
The model offers two ways to process a document. In open extraction, it identifies information from the page’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, mark absent values as null and arrange table entries row by row. Hugging Face says one checkpoint can handle different document types without fine-tuning for each type.
The company describes uses such as pulling suppliers, invoice numbers, due dates and totals from invoices, or extracting requisition details for scheduling systems. It says the model builds on DharmaOCR-LITE v1.1, a 2-billion-parameter OCR model, with a LoRA adapter trained on 43,384 examples. The examples included English documents from the public web, internally sourced Brazilian documents in Portuguese, and smaller sets of other European and Asian languages. Document types included receipts, invoices, forms, bank statements, IDs and synthetic documents.
Hugging Face reports evaluating the extractor on 546 documents and 1,638 cases, spanning open extraction, partial schemas and schemas that requested fields absent from the document. Its scoring combined field-level F1 by path, including table position, value similarity and the share of valid JSON outputs. The company says its quality results ranged from 1.8 to 9.3 points above the four tested configurations. It also reports estimated costs roughly six to 147 times lower, based on average tokens per page and providers’ published list prices.
From Scanned Pages to Business Records
Document processing often requires more than recognizing words. OCR transcribes page content; a separate step must identify which value is an invoice date, supplier or total and place it in a form another system can use. A model that returns consistent, schema-shaped data could reduce custom mapping and manual entry for organizations processing large numbers of records.
The two request modes address different needs: a schema can fit a system with fixed fields, while open extraction may help when the relevant fields are not known ahead of time. If the reported cost advantage and output quality hold on an organization’s own documents, the model could make automated extraction more practical at scale. But benchmark scores and token-based estimates do not establish the total cost or reliability of a deployed workflow. Error handling, document variation and human review can affect both.
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
Apple Watch SE 3 GPS Smartwatch, 40mm
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Hugging Face’s OCR Work
Hugging Face released DharmaOCR-LITE in April, according to the supplied material, as a model for reading varied documents such as contracts, court filings, invoices, IDs and forms. The year of that release is not specified. The new Field Extractor adds an interpretation step: instead of returning text that follows the page, it assigns values to fields and can match table entries to columns.
The company presents this as a task-specific extension of its OCR work, trained on field-level annotations across document types and languages. The supplied account does not give the publication date of the Field Extractor announcement. It also contrasts the model with general-purpose language models, which may need task-specific prompting and can cost more to run at volume; the reported comparisons, however, apply only to the configurations included in Hugging Face’s test.
“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 Deployment Gaps
The announcement summary does not provide individual model scores, the score scale or full benchmark tables, making it difficult to assess the reported quality gap. It also does not give the Field Extractor’s own price, a public endpoint, deployment requirements or a release date. The cost estimates use average input and output tokens per page, including reasoning tokens, and published provider list prices; actual costs may differ across workloads and usage arrangements.
There is no independent reproduction or evidence in the supplied material about performance on unseen document sets. Error rates by language and document type, results on low-quality scans, and the amount of human checking required are also unknown. The benchmark’s results do not establish how the model would perform in any particular organization’s workflow, especially where incorrect or missing fields have operational consequences.
Evidence Needed for Adoption
Hugging Face says users can try the model, but the supplied information does not specify where or how it can be accessed. The next evidence needed to judge its practical value includes complete benchmark results and reproducible evaluation details, followed by tests on varied real-world documents and field requirements.
Organizations considering use will also need to assess how the model handles absent or incorrect values, how sensitive documents are processed, and when a person must review the output. Until those details and independent results are available, the reported benchmark and cost advantage should be treated as company-reported results, not proof of reliable automation across business settings.
Key Questions
What does Dharma Field Extractor do?
It takes a document image and returns structured JSON, either by identifying fields on its own or filling a schema supplied by the user.
How is it different from OCR?
OCR transcribes text from a page. Field extraction adds a step that assigns values to named fields, such as identifying an invoice number or matching table entries to columns.
What benchmark results did Hugging Face report?
Hugging Face says it tested 546 documents across 1,638 cases and reported quality results 1.8 to 9.3 points above four frontier model configurations. The supplied material does not include the full scores or score scale, and the results have not been independently validated in that material.
Are the reported cost savings guaranteed?
No. The company reports estimates roughly six to 147 times lower than the tested configurations, using average token use and published list prices. The Field Extractor’s own price and the costs of review, error handling and deployment are not provided.
Can it handle fields that are missing from a document?
In schema-directed mode, the model is intended to return null for requested fields absent from the document. The supplied information does not establish how reliably it does so across different documents or real-world workflows.
Primary source: Hugging Face · via ThorstenMeyerAI.com
