Dharma Field Extractor: 2B Model Hits 0.834 F1 at 147x Lower Cost

Executive Overview

Dharma-AI has announced the launch of Dharma Field Extractor, a 2-billion-parameter model that extracts structured JSON fields from any document type. According to the official Hugging Face blog, the model outperforms Claude Opus 5 and GPT-5.6 Luna on the field-level F1 benchmark while operating at 6 to 147 times lower cost. The model is available immediately for testing and download on Hugging Face, with self-hosting on customer infrastructure.

Deep-Dive Features & Architecture

Dharma Field Extractor is built on DharmaOCR-LITE v1.1, a 2-billion-parameter model that is an optimized version of the original 3-billion-parameter model published in April in a research paper. The model was trained via supervised fine-tuning (SFT) and direct preference optimization (DPO), then a LoRA adapter was added, trained on 43,384 training examples covering both extraction modes. The data relies on three streams: multi-domain English data from the public web, real Brazilian Portuguese data, and multilingual data including European and Asian languages.

The model offers two extraction modes:

• Open Mode: The model discovers structured fields from context and layout without a predefined list and returns them as key-value pairs.
• Schema-directed Mode: The user defines the exact shape of fields with their types (string, number, boolean) using the key names their system uses, and the model returns null for missing fields instead of fabricating values. This ensures full compatibility with downstream systems without writing intermediate code.

On an SROIE invoice sample, the model extracted 18 fields with perfect accuracy, including a multi-line billing address as a single field, and returned two empty fields as null. On the FUNSD dataset, the model accurately identified a single specific box out of 12 logical Boolean fields with a nested set of 5 sub-options, and correctly read a handwritten Bates number.

Benchmark & Competitive Performance

The benchmark included 546 documents across three request scenarios (full extraction without schema, partial schema, and schema including non-existent fields), totaling 1,638 test cases. The score is calculated by multiplying field-level F1 (with path matching inside tables) by Levenshtein similarity of matched values, then by the ratio of valid JSON outputs. The model scored 0.834 versus 0.816 for Claude Opus 5 (closest competitor) and 0.741 for GPT-5.6 Luna. This represents a quality gap of 1.8 to 9.3 points in favor of Dharma Field Extractor.

Industry Impact & Enterprise Adoption

The launch signals a growing trend toward specialized small language models (SLMs) that deliver frontier-level accuracy on narrow tasks at a fraction of the cost. By self-hosting on a single L40S GPU, enterprises can process sensitive documents entirely on-premises, addressing data sovereignty and compliance requirements. The schema-directed mode eliminates the need for custom parsing code, reducing integration time and error rates. With pricing at $2.24 per GPU hour and no per-token fees, Dharma Field Extractor offers a predictable cost structure for high-volume document processing workflows such as invoice automation, contract analysis, and form digitization.

Conclusion

Dharma Field Extractor demonstrates that a 2B-parameter model can outperform much larger frontier models on structured field extraction while being dramatically cheaper to operate. Its availability on Hugging Face and self-hosted deployment options make it immediately accessible to developers and enterprises. While Arabic support is not explicitly confirmed, the multilingual training and interactive Space allow for quick validation. This release reinforces the viability of specialized SLMs for production-grade document AI.

Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis