Key points
- The Dharma Field Extractor uses two billion parameters and handles invoices, IDs and contracts without per-type fine-tuning.
- Outputs are structured JSON that follow user-defined schemas exactly, returning null for missing fields rather than hallucinating values.
- The model supports both open extraction of unknown fields and schema-directed requests for specific system columns.
The company states the model outperforms frontier large language models in this task while costing significantly less to run.
The evidence
While OCR preserves page formatting for full-document digitisation, the new model targets downstream systems that require specific values, such as invoice totals or patient names, rather than raw text.
The decision
Users can request data in two ways: an open mode that identifies structured fields from context, or a schema-directed mode that fills a predefined JSON shape. The schema mode ensures output keys match existing system columns, returning null for absent fields instead of inventing data, which simplifies integration into databases.
What to make of it
The blog post argues that using general-purpose LLMs for high-volume field extraction is expensive due to API pricing and latency. By specialising in this narrow task, the smaller model aims to provide a more efficient solution for businesses processing batches of supplier invoices, medical requisitions or compliance contracts daily.
In practice
Why it matters
- Outperforms frontier LLMs on field extraction tasks while operating at a fraction of the cost and latency.
- Returns structured JSON with typed fields and null values for missing data, enabling direct integration into downstream systems without additional mapping code.
- Supports both open extraction for unknown document structures and schema-directed extraction to match existing system keys and types exactly.
- Handles high-volume business needs for specific data points, such as invoice line items or medical requisition details, more efficiently than general-purpose APIs.
Topics: models, document processing, Hugging Face, open source, AI
Updated
The strongest doubt
Counterpoint from the sources
The post claims superiority over frontier LLMs but does not provide independent benchmarks or third-party validation of these performance figures.
Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Does the model require fine-tuning for each new document type?
No, the blog post states the same checkpoint handles all document types without per-type fine-tuning.
What happens if a requested field is not in the document?
The model returns null for missing fields instead of inventing a value, according to the post.
Connected stories
Reading receipt
- The Dharma Field Extractor uses two billion parameters and handles invoices, IDs and contracts without per-type fine-tuning.
- Outputs are structured JSON that follow user-defined schemas exactly, returning null for missing fields rather than hallucinating values.
- The model supports both open extraction of unknown fields and schema-directed requests for specific system columns.