🔍 Read the full analysis: How To Get Structured JSON Fields From A Document 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 extracts document information into JSON, either by identifying fields 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, independent validation and the model’s own pricing are not included in the supplied material.
Hugging Face has introduced Dharma Field Extractor, a 2-billion-parameter model designed to turn document images into structured JSON. The company says it performed better on its benchmark and had lower estimated operating costs than four tested frontier model configurations, though the provided information does not include full scores or independent testing.
The model supports two ways to request data. In open extraction, it identifies information from labels and page layout and returns key-value pairs. In schema-directed extraction, a user supplies field names and types; the model is intended to return those keys, use null when a requested field is absent, and arrange table contents row by row.
Hugging Face says a single checkpoint handles different document types without per-type fine-tuning. The examples it gives include extracting suppliers, invoice numbers, due dates and totals from invoices, or passing requisition details to a scheduling system. These are described as potential uses, not evidence that the model has been validated in those workflows.
The company says the extractor builds on DharmaOCR-LITE v1.1, a 2-billion-parameter OCR model, using a LoRA adapter trained on 43,384 examples. Those examples included English documents from the public web, internally sourced Portuguese documents from Brazil, and a smaller collection of other European and Asian languages. Document categories included receipts, invoices, forms, bank statements, identification documents and synthetic examples.
From Scanned Pages to Business Records
Document systems often need more than a transcript of a page: they need specific values placed into fields that other software can read. Structured JSON output could reduce custom mapping or manual data entry if it reliably identifies the right values and follows the required format. Schema-directed requests may fit established systems with fixed fields; open extraction may help when the relevant labels are not known in advance.
Hugging Face positions the model for repeated, high-volume work such as invoice processing. Its reported cost advantage could matter at that scale, but the figures are estimates, not a total workflow-cost comparison. Real costs and usefulness may depend on document quality, volume, error correction, human review and the systems used to process the output. The benchmark summary does not establish how the model would perform for any particular organization.
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
Logitech MX Keys Combo for Business Keyboard and Mouse – Graphite
EcoFlow Delta 3 Max Portable Power Station – 2048 Wh LiFePO4 Battery, 2400 W Output (3400 W X‑Boost)
As an affiliate, we earn on qualifying purchases.
How Dharma Adds Fields to OCR
OCR generally converts printed or written page content into text. Field extraction adds an interpretation step: it associates values with named fields and may organize table entries by row and column. That distinction matters because a readable text transcript is not, by itself, a record ready for an invoice system or database.
Hugging Face says it released DharmaOCR-LITE in April for reading varied documents, including contracts, court filings, invoices, identification documents and forms. The supplied material does not specify the year of that release or the publication date of the Field Extractor announcement. It describes the new model as a task-focused extension of the company’s document OCR work, trained with field-level annotations across multiple document types and 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 evaluation covered 546 documents and 1,638 cases, including open extraction, partial schemas and schemas containing fields absent from a document. Its scoring combined field-level F1 by path, including table position, with value similarity and the share of valid JSON outputs. The supplied summary reports a quality gap of 1.8 to 9.3 points against four frontier configurations, but does not give the score scale, individual scores or full results.
The company also reports estimated costs roughly six to 147 times lower than those configurations. Its calculation used average input and output tokens per page, including reasoning tokens, and providers’ published list prices. The material does not state Dharma Field Extractor’s own price. It also does not provide independent reproductions, results on unseen document collections, error rates by language or document type, or an estimate of how much human review is needed. The benchmark’s reasoning settings were low, with adaptive effort for Claude Opus 5, according to the supplied account.
Testing Outside the Benchmark
Hugging Face says users can try the model, but the supplied material does not provide a release date, public endpoint or deployment requirements. The next evidence readers should look for includes complete benchmark tables, reproducible evaluation details and tests by organizations using different document types, image quality and field requirements.
For teams considering deployment, performance is only one part of the decision. They will also need to determine how to catch and correct incorrect values, handle sensitive documents, and respond when a requested field is missing. Whether the reported quality and cost results carry over to real workloads remains unconfirmed.
Key Questions
What does Dharma Field Extractor do?
Hugging Face describes it as a 2-billion-parameter model that reads document images and returns information as structured JSON.
How do its two extraction modes differ?
Open extraction asks the model to find structured information on its own. Schema-directed extraction gives it specific field names and types to fill, with null intended for requested values that are absent.
What benchmark results did Hugging Face report?
The company says its evaluation covered 546 documents and 1,638 cases, with quality 1.8 to 9.3 points above four tested frontier configurations. The supplied information does not specify the score scale or include individual scores, so the result cannot be independently interpreted from this summary alone.
Are the reported cost savings confirmed for businesses?
No. Hugging Face reports estimates roughly six to 147 times lower than the tested configurations, calculated from average token use and published list prices. The material does not state the model’s own price or establish total costs in business workflows.
Has the model been independently validated?
The supplied material does not describe independent testing. Performance on unseen documents, error rates across document types and languages, and the amount of human review required remain unclear.
Primary source: Hugging Face · via ThorstenMeyerAI.com
