Hugging Face releases Dharma Field Extractor, a 2B-parameter model for structured document field extraction
The open-source model converts any document image into typed JSON fields, outperforming frontier LLMs on a 546-document benchmark while cutting costs by up to 147×.
1 source · cross-referenced
- A 2-billion-parameter model that extracts user-specified fields from any document type and returns structured JSON, with no per-type fine-tuning required.
- Outperforms four frontier LLMs on a 546-document benchmark with field-level F1 × Levenshtein × valid-JSON scoring, with quality gaps from 1.8 to 9.3 points.
- Costs up to 147× lower than frontier API pricing when self-hosted on a single L40S GPU ($2.24/hour), with measured token usage comparisons provided.
- Supports open requests (no schema) and schema-directed extraction with typed fields, handling nested tables and returning null for missing fields.
Dharma Field Extractor is a new 2-billion-parameter model designed to extract structured fields from any document image and return them as typed JSON, without requiring per-type fine-tuning. It handles both open requests—where the model infers fields from layout and context—and schema-directed requests, where users specify exact field names and types. In schema mode, the output strictly adheres to the requested schema, including nulls for missing fields, enabling direct integration into downstream systems.
The model is built on top of DharmaOCR-LITE v1.1, a 2B-parameter OCR-specialized checkpoint, with a LoRA adapter trained on 43,384 field-level annotated examples spanning English multi-domain, Brazilian domain (Portuguese), and multilingual data. The training mix includes receipts, invoices, forms, bank statements, ID documents, and synthetic documents to ensure multi-domain generalization.
Evaluation covers 546 documents across three scenarios: extracting all fields with no schema, a partial schema with only real fields, and a schema that includes deliberately nonexistent fields. The benchmark uses a composite metric combining field-level F1 (matched by path, including table positions), Levenshtein similarity on matched values, and the share of outputs that are valid JSON. Against four frontier LLM configurations, the model’s quality gap ranges from 1.8 points (Claude Opus 5) to 9.3 points (GPT-5.6 Luna).
Cost comparisons show the model’s self-hosted deployment on a single L40S GPU ($2.24/hour) is up to 147× cheaper than frontier API pricing, based on measured average input/output tokens per page and published list prices. Frontier models were evaluated with reasoning set to low, and Claude Opus 5 used adaptive effort. The self-hosted cost figure is illustrative and depends on hardware, volume, and latency requirements.
Generalization was tested on public datasets SROIE and FUNSD, which were not used in training. On SROIE, the model correctly extracted all 18 fields from a receipt, including multi-line addresses and blank fields returned as null. On FUNSD, it accurately identified a single checked boolean field within a nested group of five sub-options on a photocopying variance form.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79