🔍 Read the full analysis: A Simpler Way To Extract Fields From Documents Into JSON 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 turns document images into JSON, either by identifying fields or filling a user-provided schema. The company reports better results and lower estimated costs than four tested frontier-model configurations, but full scores, pricing details and independent validation were not provided.
Hugging Face has introduced Dharma Field Extractor, a 2-billion-parameter model that converts document images into structured JSON. The company says it can either identify fields on its own or fill a schema supplied by a user, and reports better benchmark quality than four tested frontier-model configurations at substantially lower estimated cost.
The extractor offers two modes. With an open request, it identifies information from labels and page layout and returns key-value pairs. With a schema-directed request, a user names the fields and their types; the model is intended to return those keys, use null when a requested value is absent, and arrange table entries row by row. Hugging Face says one checkpoint works across document types without per-type fine-tuning.
The company describes use cases such as taking supplier names, invoice numbers, due dates and totals from invoices, or sending requisition details to scheduling systems. It says the benchmark covered 546 documents and 1,638 cases spanning open extraction, partial schemas and schemas that included nonexistent fields. The reported quality gap over four tested configurations ranged from 1.8 to 9.3 points; the supplied material does not state the score scale or include individual scores.
Hugging Face also reports costs roughly six to 147 times lower than those configurations. The estimates use average input and output tokens per page, including reasoning tokens, and providers’ published list prices. The supplied information does not give the Field Extractor’s own price. The benchmark used low reasoning settings, with adaptive effort for Claude Opus 5, according to the source material.
From Scanned Pages to Business Records
Document processing often requires more than reading text. An OCR system can transcribe a page, but another step is needed to map a number to “invoice total,” identify a due date, or place table entries under the correct columns. Dharma Field Extractor is aimed at that handoff from page text to usable records.
If it reliably returns accurate values in the format a business system expects, schema-directed extraction could reduce custom mapping and manual data entry. That could matter for organizations processing many invoices, forms or statements. Open extraction may also help when the relevant fields are not known in advance, while fixed schemas suit workflows that expect consistent keys and data types.
The reported price advantage could be relevant at high volume, but it is not a measure of total operating cost or proof that a workflow can run without review. Real costs may include retries, exception handling, integration and people checking errors. The results supplied by Hugging Face do not establish how the model would perform on a particular organization’s documents.
Products Worth Considering
Anker SOLIX C300 Portable Power Station, 288Wh
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
As an affiliate, we earn on qualifying purchases.
Built on Hugging Face OCR Work
The new extractor builds on DharmaOCR-LITE v1.1, a 2-billion-parameter model specialized for optical character recognition. Hugging Face says it trained a LoRA adapter for field extraction using 43,384 examples. Those examples included English documents from the public web, internally sourced Brazilian documents in Portuguese, and a smaller set in other European and Asian languages.
The reported document categories include receipts, invoices, forms, bank statements, IDs and synthetic documents. Hugging Face says it released DharmaOCR-LITE in April for reading varied documents, including contracts and court filings. The source material does not identify the year of that release or give a publication date for the Field Extractor announcement.
OCR generally provides text in page order. Field extraction adds an interpretation step by assigning values to named fields and, for tables, relating entries to rows and columns. Hugging Face presents the new model as a task-specific extension of its OCR work, rather than simply a text transcription tool.
Benchmark Results Need More Detail
The results are company-reported benchmark findings; the supplied material does not include complete tables, individual model scores, or independent reproduction. Although it gives a quality-gap range of 1.8 to 9.3 points, it does not identify the scale. The cost comparison is also an estimate based on average token use and published list prices, and the extractor’s own price is not provided.
It remains unclear how the model performs on unseen documents, how accuracy varies by language or document type, and how often a person would need to review or correct its output. The benchmark’s size and categories do not establish performance on every organization’s documents, which may differ in image quality, layout and field requirements. The material also does not detail deployment requirements, data handling for sensitive documents, or how the system behaves when information is ambiguous or absent beyond its stated null behavior for missing fields.
What Teams Need to Test
Hugging Face says users can try the model, but the supplied material does not specify a release date, public endpoint or deployment requirements. Useful next evidence would include complete benchmark results, reproducible evaluation methods, and tests on varied documents outside the company’s benchmark.
Organizations considering the model would need to test it against their own document mix and measure both extraction errors and the work required to handle them. They would also need to decide how to review uncertain outputs, protect sensitive records, and route missing or inconsistent fields. Until those details are available, the reported results indicate a promising task-specific tool, not proof of reliable automation across business settings.
Key Questions
What does Dharma Field Extractor do?
It turns document images into structured JSON, either by identifying fields itself or by filling fields specified in a user-provided schema.
How does it differ from OCR?
OCR transcribes text from a page. Field extraction adds a mapping step, assigning values to named fields and, for tables, organizing entries into rows and columns.
What performance and cost results did Hugging Face report?
Hugging Face says the model’s benchmark quality exceeded four tested frontier-model configurations by 1.8 to 9.3 points and estimated its costs at roughly six to 147 times lower. The supplied material does not give the score scale, full scores or the extractor’s own price.
Has the model been independently validated?
No independent validation is documented in the supplied material. The benchmark results and cost estimates are attributed to Hugging Face.
What should organizations check before using it?
Teams should test accuracy on their own document types and languages, estimate the need for human review, and assess integration, error handling and treatment of sensitive documents.
Primary source: Hugging Face · via ThorstenMeyerAI.com
