🔍 Read the full analysis: Turn Document Content Into JSON Fields With Dharma on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Hugging Face has introduced Dharma Field Extractor, a 2-billion-parameter model that converts document images into JSON, either by finding fields itself or following a supplied schema. The company reports higher benchmark quality and lower estimated costs than four tested frontier-model configurations, but the supplied material lacks full scores and independent validation.
Hugging Face has introduced Dharma Field Extractor, a 2-billion-parameter model that turns document images into structured JSON. It can identify fields on its own or fill a schema provided by a user; the company says it outperformed four tested frontier-model configurations on its benchmark at substantially lower estimated cost, though full results and independent validation were not included in the supplied material.
The extractor 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 mode, users specify field names and types; the model is intended to return those keys, use null when a requested value is absent, and organize table entries row by row. Hugging Face says a single checkpoint handles different document types without fine-tuning for each one.
The company describes uses such as pulling supplier names, invoice numbers, due dates and totals from invoices, or passing requisition details into 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. Those examples included English documents from the public web, internally sourced Brazilian Portuguese documents, and smaller sets of other European and Asian languages. The document categories included receipts, invoices, forms, bank statements, identification documents and synthetic materials.
Hugging Face says its evaluation covered 546 documents and 1,638 cases, including open extraction, partial schemas and schemas that requested fields not present in the document. It reports a quality gap of 1.8 to 9.3 points over the four tested configurations and estimated costs roughly six to 147 times lower. The provided summary does not identify the score scale, list individual model results or give the extractor’s own price.
From Scanned Pages to Business Records
Document extraction addresses the step between reading a page and entering its contents into software. OCR can transcribe words, but a business system often needs a value assigned to a specific field, such as an invoice total or account number. Dharma Field Extractor is designed to produce those structured, machine-readable records directly, potentially reducing custom mapping and manual data entry when it performs reliably.
The two request modes serve different needs: a fixed schema can match an existing database or form, while open extraction may help when a document’s fields are not known in advance. The reported cost advantage could matter to organizations handling many pages. But the company’s estimates use average tokens per page and published list prices, and do not establish the total cost of a deployed workflow, which also depends on document quality, error correction, review and volume.
For businesses, the practical question is not only whether a model returns valid JSON, but whether the values are correct often enough to reduce work without introducing costly mistakes. The available benchmark summary offers an initial comparison, not evidence that the model will work equally well across a particular organization’s documents or languages.
Products Worth Considering
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)
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
As an affiliate, we earn on qualifying purchases.
Building on DharmaOCR-LITE
Hugging Face released DharmaOCR-LITE in April, according to the supplied material, for reading varied documents including contracts, court filings, invoices, identification documents and forms. The material does not specify the year of that release. The new Field Extractor extends that document-reading work: OCR follows the text on a page, while extraction interprets it by assigning values to named fields and matching table entries with columns.
The company says it trained a LoRA adapter on field-level annotations spanning multiple document types and languages. This is presented as a task-specific approach to extracting information rather than a general-purpose language model prompted to handle each document. Hugging Face’s announcement frames the model around both flexible extraction and output that conforms to a user’s requested schema.
Hugging Face’s description says users can try the model, but the supplied material does not provide an announcement date, a public endpoint or deployment requirements. It also does not include a link to full benchmark tables or a detailed account of how the examples were split between training and evaluation.
“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
The reported results are company-reported benchmark claims; the supplied material does not include the full score tables, individual results for the four comparison configurations or the scale behind the 1.8-to-9.3-point quality gap. It also does not provide an independent reproduction or performance results from organizations using their own documents. The figures therefore do not establish how the model will perform in a specific workflow.
The cost comparison is also an estimate, not a complete operating-cost calculation. Hugging Face says it used average input and output tokens per page, including reasoning tokens, and providers’ published list prices. The summary does not state the Field Extractor’s own price, nor does it account for the cost of human review, correcting errors or managing sensitive documents.
Other open questions include error rates by language and document category, performance on unseen or low-quality documents, how often absent fields are handled correctly, and how much human checking is needed before results can be used. The material also does not specify a release date or deployment requirements.
Testing Against Real Document Workflows
The most informative next evidence would be complete benchmark results and reproducible evaluation details, followed by independent tests across different document sets, languages and image quality levels. Organizations considering the model will need to compare it with their current process using their own fields and error tolerances, rather than relying only on aggregate benchmark results.
Before deployment, teams will also need to decide how to handle incorrect values, missing fields and sensitive documents, and where human review is required. The supplied material does not announce a next benchmark release, product milestone or deployment schedule, so those developments remain unconfirmed.
Key Questions
What does Dharma Field Extractor do?
It converts document images into structured JSON, either identifying fields itself or filling field names and types supplied by the user.
How does it differ from OCR?
OCR transcribes text from a page. Dharma Field Extractor is designed to go further by assigning information to fields and organizing table values into structured output.
How strong were its benchmark results?
Hugging Face reports a 1.8-to-9.3-point quality advantage over four tested frontier-model configurations, but the supplied material does not give the score scale or full individual results. The company also reports estimated costs roughly six to 147 times lower for those comparisons.
Have the results been independently verified?
The supplied material does not report independent validation. It also does not provide results from organizations testing the model on their own document collections.
What remains unknown before adoption?
Organizations still need evidence about accuracy on their documents, error rates by language and document type, human-review needs, the model’s own price, and deployment requirements. The supplied material does not answer those questions.
Primary source: Hugging Face · via ThorstenMeyerAI.com
