TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Dharma AI has introduced Dharma Field Extractor, a 2-billion-parameter model that turns document images into JSON fields, either by finding fields on its own or following a supplied schema. The company reports that it outperformed four tested frontier-model configurations on its benchmark and cost less to run, but the results come from Dharma AI’s own evaluation.
Dharma AI has introduced Dharma Field Extractor, a 2-billion-parameter model that converts document images into structured JSON, either identifying fields itself or filling a schema supplied by the user. The company says the model outperformed four tested frontier-model configurations on its benchmark and ran at a lower estimated cost, though the results are from the company’s evaluation rather than an independently reported comparison.
The model is designed for workflows that need selected values rather than a full text transcription: for example, invoice totals and line items, details from identity documents, or fields from medical requisitions. Dharma AI says users can request an open extraction, in which the model returns fields it identifies from the document, or provide a schema with named fields and data types. The company says schema-directed outputs use the requested keys, add no extra fields and return null when a requested value is absent.
Dharma AI says one checkpoint handles multiple document types without per-type fine-tuning. Its reported training setup uses a LoRA adapter trained on 43,384 examples, including English documents collected from the public web, internally sourced Brazilian documents in Portuguese, and a smaller multilingual set covering other European and Asian languages. The source lists examples such as receipts, commercial invoices, forms, bank statements, identity documents and synthetic documents.
The company evaluated the model on 546 documents and 1,638 cases spanning three scenarios: extracting all fields without a schema, completing a schema containing real fields, and completing one that also includes nonexistent fields. Dharma AI says its score combines field-level matching, similarity of matched values and the share of outputs that are valid JSON. It reports a quality advantage of 1.8 to 9.3 points over four general-purpose configurations and estimated costs roughly six to 147 times lower, depending on the comparison. These are company-reported benchmark and pricing results.
From OCR Text to Usable Records
Document processing often has two distinct steps: reading text from a page and assigning that text to fields a business system can use. DharmaOCR-LITE, the company’s earlier OCR model, was aimed at returning text faithful to a document’s formatting. Field Extractor targets the next step, producing structured values rather than a page transcription.
If the reported performance holds in practical deployments, a model that can map document contents directly into existing schemas could reduce manual review or custom extraction code in high-volume workflows. Its ability to represent tables as ordered rows and to mark missing requested fields as null may also help downstream systems handle invoices or forms consistently. But benchmark scores alone do not establish how often the model makes consequential mistakes on a particular organization’s documents; users would need to test it on their own data and review process.
The cost comparison could matter for organizations that process many pages, since the company positions specialized extraction as less expensive than using general-purpose APIs. The reported range varies by model and depends on estimated input and output tokens and published list prices. It should not be read as a guaranteed saving for every workload.
Products Worth Considering
EcoFlow DELTA 3 Portable Power Station (1024Wh, 1800W)
Bose SoundLink Plus Portable Bluetooth Speaker, Black
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
As an affiliate, we earn on qualifying purchases.
Built on DharmaOCR-LITE
Dharma AI describes Field Extractor as an extension of DharmaOCR-LITE, which it says it released in April. The earlier model focused on OCR across document types including contracts, court filings, invoices, identity documents and forms. The company’s rationale for the new release is that transcription alone does not identify which extracted string is, for example, an invoice number, due date or total.
According to the report, the Field Extractor base is DharmaOCR-LITE v1.1, a 2-billion-parameter version optimized from the earlier model and adapted for OCR with supervised fine-tuning and direct preference optimization. Dharma AI then trained a LoRA adapter for field extraction. The supplied report describes the model, training data and evaluation, but does not provide an independent assessment or customer deployment results.
“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.”
— Dharma AI, in its Hugging Face report
Limits of the Reported Evaluation
The supplied report does not name the 546 benchmark documents in enough detail to establish how closely they represent any particular customer’s documents, nor does it describe an independent replication. It also does not provide the full per-model score table or the exact results for every request scenario in the material provided. The company’s claim that Field Extractor outperformed every tested frontier configuration applies to the four configurations it says it evaluated, not to all available models.
The cost estimates are based, according to Dharma AI, on measured average input and output tokens per page, including reasoning tokens, and providers’ published list prices. The provided material says frontier models ran with reasoning set to low and that Claude Opus 5 used a different setting, but the relevant sentence is incomplete. The exact model versions, full price calculations and conditions needed to reproduce all comparisons are therefore not fully clear from the supplied text.
It is also not clear from the report how the model performs on documents outside its training mix, how it handles poor-quality images or unusual layouts, or what human review is recommended for sensitive records. The company’s statement that missing fields return null is a product claim; the source does not provide a separate error-rate analysis for false or missed extractions.
Testing on Real Document Workflows
Dharma AI says the model can be tried through the link in its Hugging Face report. The immediate next step for prospective users is to evaluate it on representative documents and schemas, checking both extraction accuracy and valid JSON output before connecting it to production systems. Testing should include absent fields, tables and the document variations that matter to each workflow.
The source does not announce a release schedule, independent benchmark, or named customer deployment. Further detail on model access, deployment options, full benchmark results and performance across languages would help clarify how the reported results translate beyond the company’s evaluation.
Key Questions
What does Dharma Field Extractor do?
It takes a document image and returns fields as JSON. Users can ask it to identify fields it finds or provide a schema specifying the desired keys and types.
How does it differ from OCR?
OCR transcribes text from a page. Field Extractor is intended to organize selected content into named fields and structured tables for downstream systems, rather than returning only the document’s text.
What benchmark results did Dharma AI report?
The company says it tested 546 documents across 1,638 cases and scored field matching, matched-value similarity and valid JSON output. It reports a 1.8-to-9.3-point quality advantage over four tested general-purpose configurations; the source does not establish an independent replication.
Does the model work across document types and languages?
Dharma AI says its training included receipts, invoices, forms, bank statements, identity documents and synthetic documents, with English, Portuguese and a smaller multilingual component. The report does not provide a separate performance breakdown for each document type or language.
What happens when a requested field is missing?
Dharma AI says schema-directed extraction returns null for a field that is not present, rather than filling it with an invented value. The supplied report does not give a separate error rate for this behavior.
Source: huggingface.co
