Dharma Field Extractor Beats Frontier LLMs on JSON Document Extraction
Dharma-AI's 2B Field Extractor turns documents into schema-ready JSON, scoring 0.834 and costing 6 to 147 times less than tested frontier LLMs in benchmarks.
Summary
On November 4, 2026, Dharma-AI introduced Field Extractor, a 2-billion-parameter model that converts document images into typed, nested JSON without per-document-type fine-tuning. It supports open extraction or an exact schema, preserves requested keys and types, fills tables cell by cell, adds no unspecified fields and returns null rather than inventing missing values. The model builds on DharmaOCR-LITE v1.1, already specialized through supervised fine-tuning and DPO, using a LoRA adapter trained on 43,384 field-annotated examples spanning English public-web documents, internal Brazilian Portuguese documents and a smaller multilingual European and Asian set.
Dharma-AI tested 546 documents across three scenarios, producing 1,638 cases covering open extraction, real-field schemas and schemas containing nonexistent fields. Its metric combines field-level F1, path and table-position matching, Levenshtein value similarity and valid JSON frequency. Field Extractor scored 0.834, ahead of four frontier configurations run with low reasoning: Claude Opus 5 scored 0.816 using adaptive effort, while GPT-5.6 Luna scored 0.741. The quality advantage ranged from 1.8 to 9.3 points.
Estimated extraction cost was roughly 6 times below GPT-5.6 Luna and 147 times below GPT-5.6 Sol, based on one self-hosted L40S costing $2.24 hourly, although deployment economics vary by hardware, volume and latency. Held-out SROIE and FUNSD tests correctly recovered 18 receipt fields, 12 checklist booleans, a five-option nested group, handwriting and blank fields. The release extends April's 3-billion-parameter DharmaOCR-LITE OCR work into downstream field extraction.
Positives
- Field Extractor scored 0.834, beating Claude Opus 5 at 0.816 and GPT-5.6 Luna at 0.741.
- Estimated extraction costs were roughly 6 to 147 times lower than the compared frontier APIs.
- All 18 fields in a held-out SROIE receipt example were recovered correctly, including two blanks returned as null.
- A held-out FUNSD form produced all 12 boolean values correctly, plus its handwritten Bates ID and nested options.
- One checkpoint supports invoices, IDs, forms, bank statements and other documents without per-type fine-tuning.
Risks & concerns
- The 0.834 result comes from Dharma-AI's own 546-document benchmark rather than an independent evaluation.
- Self-hosted costs are not sales prices and can change with hardware, workload volume and latency requirements.
- Only four frontier configurations were compared, all using low reasoning, although Claude Opus 5 used adaptive effort.
- European and Asian languages represented a smaller training share than English and Brazilian Portuguese.