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 a 2-billion-parameter Field Extractor that turns document images into structured JSON, either by identifying fields automatically or following a user-defined schema. The company reports that it outperformed four tested general-purpose models on its benchmark and cost less to run; the benchmark and pricing comparisons are from Dharma AI, and independent validation was not provided in the source material.
Dharma AI has introduced the Dharma Field Extractor, a 2-billion-parameter model that converts document images into structured JSON fields, either finding fields on its own or filling a schema specified by the user. The company says the model outperformed four general-purpose language-model configurations in its benchmark while costing less to run, a claim relevant to organizations processing documents at high volume.
The extractor accepts a document image and offers two modes. In open extraction, it identifies structured information from the document’s text and layout, such as labeled values and table rows, then returns key-value pairs. In schema-directed extraction, a user supplies field names and types—such as strings, numbers or booleans—and the model returns those keys in the requested structure. Dharma says fields that are absent are returned as null, rather than filled with invented values, and schema-based responses contain no extra fields.
The intended use is to pass selected information into another system without first turning the whole page into text and writing a separate mapping step. Examples in the announcement include extracting supplier, invoice number, due date and total from invoices; collecting patient and appointment details from medical requisitions; and pulling parties, terms or penalties from contracts. Tables can be returned as ordered lists of rows, with values associated with their columns.
Dharma says the system is built on DharmaOCR-LITE v1.1, a 2-billion-parameter OCR-specialized model, with a LoRA adapter trained on 43,384 examples. The training data spans document categories including receipts, invoices, forms, bank statements and identification documents. The company describes the data as including public-web English documents, internally sourced Brazilian documents in Portuguese, and a smaller multilingual collection covering other European and Asian languages.
From OCR Text to Usable Records
OCR can recover the words on a page, but it does not by itself tell a business system which number is an invoice total or which date is a payment deadline. That gap often requires additional software or human review. A model that returns typed, schema-matched fields could reduce that extra step for organizations that process many forms, invoices or records.
The practical case depends on more than a strong benchmark score. Businesses would need the extracted values to be accurate enough for their workflows, the output to integrate with existing systems, and the price and speed to hold up at their own volumes. The null response for missing fields and the option to constrain output to an existing schema are designed to address common integration needs, but the announcement does not establish how reliably those features perform across real-world deployments.
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
EcoFlow Delta 3 Portable Power Station – 1024 Wh, 1800 W
As an affiliate, we earn on qualifying purchases.
Dharma’s OCR Model Foundation
The extractor follows Dharma AI’s earlier release of DharmaOCR-LITE, which the company presented as a specialized system for reading varied documents, including contracts, court filings, invoices, IDs and forms. OCR returns the text rather than assigning each item to a business field. The new product is intended to handle that next step: turning page content into a requested data structure.
Dharma frames the project as an application of task specialization: a smaller model trained for a defined document task may compete with general-purpose models while costing less. The company’s new comparison covers 546 documents in three scenarios—extracting all fields, filling a schema containing fields present in the document, and handling schemas that include nonexistent fields—for 1,638 test cases in total.
The reported scoring combines field-level F1, value similarity for matched fields and the share of responses that are valid JSON. Dharma says its four tested general-purpose configurations scored between 1.8 and 9.3 points below the extractor, and that the extractor’s estimated costs were roughly six to 147 times lower. These are company-reported results, not an independent 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.”
— Dharma AI, in its Hugging Face announcement
Benchmark Scope and Deployment Questions
The published results are presented by Dharma AI, which developed the model. The source material does not include an independent replication, detailed per-document-category results, or enough information here to assess how the benchmark’s documents compare with the varied and sometimes poor-quality records businesses process in production. The reported quality margin also depends on the company’s chosen test set, scoring method and model configurations.
The cost comparison is an estimate based on measured average input and output tokens per page and providers’ published list prices, according to Dharma. The announcement notes that reasoning tokens were included and that frontier models ran with reasoning set to low, but the available source text ends before completing its description of the Claude configuration. Actual costs may differ with document length, usage patterns, provider pricing and deployment setup.
Other practical details remain unclear from the supplied announcement, including processing speed, throughput under load, data retention and privacy controls, hosting options, and performance on documents outside the listed languages and categories. The model’s ability to return valid JSON does not by itself establish that every extracted value is correct.
Testing the Extractor in Workflows
Dharma says users can try the model, but the source material does not specify a release schedule, service limits or a roadmap for future versions. The next meaningful test is how it performs on documents and schemas that prospective users handle in practice, including checks for incorrect values, missing fields and table-row alignment.
Organizations considering the tool will also need to confirm its integration and data-handling terms, then compare its accuracy, latency and total cost with existing OCR, manual review and general-purpose model workflows. Until independent tests or deployment evidence are available, the benchmark figures should be treated as Dharma’s reported results, not a settled measure of performance across document-processing tasks.
Key Questions
What does Dharma Field Extractor do?
It takes a document image and returns information as structured JSON. Users can ask it to identify fields in the document or provide a schema specifying the fields and data types they need.
What happens when a requested field is missing?
Dharma says schema-directed extraction returns null for a requested field that is not present, rather than inventing a value.
How was the model trained?
Dharma says it added a LoRA adapter to its DharmaOCR-LITE v1.1 base model, training on 43,384 examples across English, Brazilian Portuguese and a smaller set of other languages. The examples cover multiple document types.
Are the benchmark results independently verified?
Not in the supplied source material. The benchmark and cost comparisons are reported by Dharma AI; no independent replication is described there.
How much cheaper is it than general-purpose models?
Dharma reports costs roughly six to 147 times lower than the four tested configurations, based on measured average tokens per page and published list prices. The comparison is an estimate, and actual costs can vary by workload and pricing.
Source: huggingface.co
