Source time Edition date
ToolsHugging Face
Dharma Field Extractor: Any Document In, Structured Fields in JSON Out
Dharma Field Extractor：任意文档输入，JSON 结构化字段输出
Dharma Field Extractor promises any-document-in, structured-JSON-out, but teams must decide how much accuracy they will trade for that convenience on messy real-world scans.
Judgment
Teams evaluating Dharma Field Extractor must choose between open extraction, which discovers fields without a predefined schema, and schema-directed extraction, which locks keys and types for downstream systems but returns null when a requested field is absent.
Basis
The Dharma-AI team says the model returns typed JSON from document images, with no per-type fine-tuning. Mechanically, open mode infers key-value pairs from layout and context, while schema-directed mode fills exact keys and types and returns null for absent fields. It is built on DharmaOCR-LITE v1.1, a 2-billion-parameter base, with a LoRA adapter trained on 43,384 examples across English multi-domain, Brazilian, and multilingual streams. The benchmark covers 546 documents across three request scenarios, for 1,638 total cases.
Unclear The source does not give per-field accuracy, latency, or cost figures for open versus schema-directed extraction.
Next Check the benchmark tables for open versus schema-directed error rates on messy scans before setting a default extraction mode.