New LoRA for Qwen3-VL-8B Adds Bounding Box Coordinates to Document Extraction

A new LoRA fine-tune for Qwen3-VL-8B-Instruct enables document extraction that includes precise location data for every value. By providing page images and a list of desired fields, the model returns each value alongside its original printed text, page number, and a bounding box, allowing for immediate verification or highlighting in review interfaces.

Key Capabilities

• Verifiable Extraction: Every field is linked to its source location via coordinates, reducing hallucination risks.
• Structured Data Handling: Supports nested fields and line items while normalizing dates and numbers.
• Null Safety: Returns null for missing fields instead of guessing, ensuring data integrity.
• No OCR Dependency: Operates directly on images without requiring separate OCR pipelines or per-document-type training.

Performance Metrics

The fine-tuned model demonstrates a significant improvement in spatial accuracy over the base model. Correct coordinate retrieval increased from 41% to 74%, making it suitable for applications requiring high-confidence document parsing.

The model is available on Hugging Face as sreeketh/Qwen3-VL-8B-DocBBox-LoRA.