Dharma Field Extractor: Any Document In, Structured Fields in JSON Out
Web page · https://huggingface.co/blog/Dharma-AI/dharma-field-extractor
First seen by Oriel on .
AI summary
Dharma Field Extractor is a 2-billion-parameter document model that returns requested fields as structured JSON and outperformed four tested frontier LLM configurations at substantially lower self-hosted cost.
- It supports open extraction of all detected fields or schema-directed extraction with exact keys, requested types, nested tables, and nulls for missing values.
- The model builds on DharmaOCR-LITE v1.1 with a LoRA adapter trained on 43,384 field-annotated examples spanning English, Brazilian Portuguese, and additional European and Asian languages.
- The benchmark used 546 documents and 1,638 cases across open, partial-schema, and nonexistent-field scenarios, scoring path-aware field F1, value similarity, and valid JSON.
- It scored 0.834, versus 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna, with estimated costs 6 to 147 times lower.
- Self-hosting costs were calculated from GPU time on one L40S at $2.24 hourly; actual deployment cost varies with hardware, volume, and latency requirements.
#document ai #information extraction #small language models #ocr
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/Dharma-AI/dharma-field-extractor | https://huggingface.co/blog/Dharma-AI/dharma-field-extractor Read discussion |