Dharma-AI Launches Field Extractor, a 2B Model That Turns Documents Into Schema-Compliant JSON
The 2B document-AI model returns structured JSON with missing fields as null instead of hallucinated values, marking document field extraction as its own product category.
Dharma-AI has announced Field Extractor, built on its DharmaOCR-LITE v1.1 base model (2B parameters, refined with SFT and DPO) plus a LoRA adapter trained on 43,384 examples. The model accepts a document image and operates in two modes: open mode, where the model detects fields from the layout itself, and schema-directed mode, where the client defines typed fields (string, number, boolean) with its own key names.
The output is structured JSON that mirrors the document's structure, with invoice line items returned as ordered lists. Missing fields are returned as null rather than fabricated values — a direct answer to the hallucination problem that has kept generic LLMs out of financial document processing.
The vendor claims it outperforms all frontier LLMs it tested on this task at lower cost and latency, though all accuracy figures are self-reported and the weights and license have not yet been published (as of November 7, only the older DharmaOCR-LITE from March 2026 is on Hugging Face). The release confirms a broader shift this week: value in document AI is moving from faithful page parsing toward programmable, structured data contracts between systems.