Have you ever spent hours parsing invoices, medical forms, or identity cards only to end up writing custom code that pulls the right numbers out of the mess? That frustration has a new answer in the form of a 2‑billion‑parameter field extraction model released by Dharma AI. Dubbed the Dharma Field Extractor, it takes a document image and produces structured JSON, with every field typed and nested exactly how your downstream system expects. Whether you’re ingesting a stack of supplier invoices every morning or feeding medical requisitions into a scheduling engine, this tool can turn the raw paper into a clean data payload in seconds, all without the need for expensive, high‑latency language‑model APIs.
The Field Extractor is built on top of the earlier DharmaOCR‑LITE, a lightweight OCR system that already outperformed larger, generic models on a variety of document types. The new model adds a “field finder” layer trained with a LoRA adapter on more than forty‑three thousand annotated examples. These examples come from a deliberately diverse mix of sources: English documents scraped from the public web, real Brazilian paperwork in Portuguese, and a smaller set of other European and Asian languages. The training set includes everything from commercial invoices and bank statements to synthetic receipts and ID cards, giving the model a broad sense of what structured data looks like across many document formats.
The Field Extractor offers two modes of operation. In the Open mode, you send the image without any schema and the model hunts for any recognizable field, returning each as a key‑value pair. It can detect labels followed by values, table headers, and even nested structures, all inferred from layout and context. The Schema‑Directed mode is even more powerful for production pipelines. You provide the exact shape you need – field names, data types, and table layout – and the model fills in those slots. If a requested field is missing from the document, it simply returns null rather than inventing a value, which keeps downstream logic simple and reliable.
Because the output is JSON, integration is straightforward. You can feed the result directly into your database, a data lake, or another microservice. The model’s precision is impressive: on tests it outperformed every frontier LLM the team evaluated, all while being cheaper to run. That makes it an attractive option for high‑volume use cases, where the cost of calling a large language model for every document would quickly add up.
To try it out, Dharma AI hosts a demo portal you can log into from your browser. There you can upload a sample document and experiment with both Open and Schema modes. The user interface is simple: you choose your mode, optionally upload a JSON schema, and hit submit. The model returns a JSON blob that you can inspect in real time. The demo also shows a side‑by‑side view of the original image and the extracted fields, so you can verify that the model is correctly interpreting the layout.
One of the key advantages of this approach is that it removes the need for domain‑specific fine‑tuning. The same checkpoint can handle invoices, forms, IDs, and even multilingual documents. The training data’s diversity and the LoRA adapter’s flexibility allow the model to generalize well, which is a huge win over more brittle, task‑specific solutions.
If you’re looking to streamline data extraction in your organization, the Dharma Field Extractor offers a compelling mix of speed, cost, and accuracy. It’s a reminder that sometimes the best way to get ahead is to build a specialized tool rather than rely on generic, one‑size‑fits‑all models. Source: Hugging Face.