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 the Dharma Field Extractor, a 2-billion-parameter model that turns document images into structured JSON, either by identifying fields or following a supplied schema. The company reports that it scored higher than four tested frontier-model configurations on its benchmark and cost less to run, but the supplied report does not include the full results table or independent verification.
Dharma AI has introduced the Dharma Field Extractor, a 2-billion-parameter model designed to turn document images into structured JSON rather than simply transcribing their text. The company says the model can either identify available fields on its own or fill a supplied schema, and that it outperformed four tested frontier-model configurations on its benchmark at lower estimated cost.
The extractor supports two request modes. In open extraction, no field list is provided: the model uses document text and layout to identify apparent labels, values and table entries. In schema-directed extraction, users supply the desired keys and data types, such as strings, numbers or booleans. Dharma AI says that mode returns the requested keys without extra fields, and uses null when a requested value is absent.
The intended output is not a page transcript but data ready for another system. The company’s examples include returning invoice line items as ordered lists, extracting identity-document details under database-compatible keys, and collecting fields from medical requisitions or contracts. One model checkpoint is intended to handle these varied document types without per-type fine-tuning, according to the report.
Dharma AI says the extractor is built on DharmaOCR-LITE v1.1, a 2-billion-parameter OCR-specialized model, with a LoRA adapter trained on 43,384 examples covering both request modes. The training data includes public-web documents in English, internally sourced Brazilian documents in Portuguese, and a smaller multilingual set spanning other European and Asian languages. Document categories include invoices, receipts, forms, bank statements, identity documents and synthetic examples.
From Document Images to System-Ready Data
The product targets a common gap in document processing: OCR can recover what a page says, but does not by itself place each value into the correct database field. A business that receives invoices, for example, may need supplier names, invoice numbers, due dates and totals in separate columns. Extracting those fields can reduce the need for a separate mapping step between transcription and data entry, if the model performs reliably on the documents a business handles.
Cost and speed may matter for organizations processing documents in large volumes. Dharma AI says its benchmark showed a quality advantage over the four tested general-purpose configurations, with the smallest reported gap at 1.8 points and the largest at 9.3 points. It also reports costs from roughly six to 147 times lower, depending on the comparison. Those figures are the company’s results, not an independent assessment, and the report describes pricing calculations based on measured average page tokens and providers’ published list prices.
The model’s schema mode could be useful where outputs must fit an existing software pipeline. But a correctly shaped JSON response is not, by itself, proof that every value is accurate. Organizations would still need to evaluate extraction errors, missing fields and document variation before relying on the output for financial, medical or compliance workflows.
Products Worth Considering
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
Logitech MX Keys Combo for Business Keyboard and Mouse – Graphite
As an affiliate, we earn on qualifying purchases.
How Dharma Built on Its OCR Model
The announcement follows Dharma AI’s earlier DharmaOCR-LITE release, described in the report as a model specialized for recognizing text across documents such as contracts, court filings, invoices, identity documents and forms. OCR is suited to preserving a page’s text for a person or system to review. The Field Extractor addresses a different task: turning selected content into named, typed values.
The company says the benchmark covers 546 documents and three request scenarios, for 1,638 cases in total. These test extracting all fields without a schema, filling a partial schema of fields present in the document, and handling schemas that include deliberately nonexistent fields. Dharma AI scores results using field-level F1, value similarity for matched fields and the share of outputs that are valid JSON. Table matches account for their position and path within the output.
The report says four general-purpose frontier configurations were tested, with reasoning set to low; it also notes a different reasoning setup for one configuration. Cost estimates use measured average input and output tokens per page, including reasoning tokens, and each provider’s published list pricing. The supplied source material does not include the full benchmark table or all configuration details.
“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 announcement
Benchmark Scope and Real-World Accuracy
The benchmark findings are reported by Dharma AI; the source material provides no independent evaluation. It does not show the full score table, detailed results by document category or language, or enough information to assess how representative the test set is of a particular organization’s paperwork. The reported quality gaps are given in points, but the supplied excerpt does not specify the underlying scores.
It is also unclear from the material how the extractor performs on low-quality scans, unusual layouts, handwriting, or documents with ambiguous or conflicting values. The report describes training across several languages, but does not provide separate accuracy results for each. The meaning of “outperformed” should be read within the stated benchmark, rather than as proof that the model will beat every general-purpose model on every document task.
The cost comparisons are estimates based on the tested configurations, token use and published list prices. Actual costs can vary with provider pricing, usage patterns and implementation. The source excerpt also does not specify deployment requirements, latency measurements, data-retention terms or whether the product is available for production use beyond the invitation to try it.
Testing Beyond the Published Benchmark
Dharma AI’s announcement invites readers to try the model, but the supplied report does not give a dated release schedule or describe a next benchmark milestone. The practical next step for prospective users is to test it against representative documents and compare extracted values with verified records, paying particular attention to tables, absent fields and non-English inputs.
Further evaluation would help establish whether the reported advantage holds across different document types and operating conditions. Relevant details would include category- and language-level scores, error rates for individual fields, latency, deployment options and the full benchmark results. Until those details are available, the company’s performance and cost figures remain claims based on its own stated test and pricing method.
Key Questions
What does the Dharma Field Extractor do?
It takes a document image and returns selected content as structured JSON. Users can ask it to identify fields it finds or provide a schema specifying the keys and types they need.
How does schema-directed extraction work?
The user supplies the required field names and data types. Dharma AI says the model returns those keys without additional fields and uses null when a requested field is not present in the document.
What documents and languages were included in training?
The company reports training on 43,384 examples covering invoices, receipts, forms, bank statements, identity documents and other material. The data includes English and Portuguese documents, plus a smaller multilingual portion covering other European and Asian languages.
Has the model been independently shown to outperform frontier models?
Not in the supplied source material. Dharma AI reports results against four general-purpose model configurations on its own benchmark; the excerpt does not provide independent verification or the full results table.
What should businesses verify before using it?
They should test extraction accuracy on their own document types, check how errors and missing values are handled, and review operational factors such as latency, cost and deployment terms. Valid JSON does not guarantee that the extracted values are correct.
Source: huggingface.co
