Zappandy/dukaan-saathi-receipt-lora
Zappandy/dukaan-saathi-receipt-lora is a 3.2 billion parameter Llama-3.2-3B-Instruct fine-tune developed by Zappandy, specifically optimized for structured receipt parsing in Indian kirana (convenience) store workflows. This model excels at extracting line items, quantities, prices, and supplier information from noisy, real-world receipt OCR text, including handwritten bills and informal tally notes. It is designed to convert diverse receipt formats into a structured JSON output, making it highly specialized for inventory management within its target domain. The model leverages a 32768 token context length to process detailed receipt information.
Loading preview...
Overview
This model, Zappandy/dukaan-saathi-receipt-lora, is a fine-tuned Llama-3.2-3B-Instruct variant, specifically developed for structured receipt parsing within the context of Indian kirana (convenience) store operations. It is an integral component of the Dukaan Saathi inventory copilot demo.
Key Capabilities
- Structured Receipt Parsing: Converts noisy OCR text from supplier receipts into a structured JSON object.
- Handles Diverse Formats: Designed to process messy, real-world receipts, including handwritten bills, printed tax invoices, and informal tally notes.
- Extracts Key Data: Identifies and extracts line items, quantities, prices, and supplier information.
Training and Limitations
- Focused Training: Trained on a small dataset of 6 hand-authored examples and 22 Modal LLM-generated synthetic examples, totaling 28 examples. This training prioritizes format consistency for specific receipt styles.
- Specialized Use: Due to its limited training data, the model may overfit to known receipt styles (e.g., Mahalakshmi Marketing, Sri Venkateshwara Marketing, Brundavan Buns) and is not a general-purpose receipt parser.
- Human Oversight: Requires owner approval before any inventory write, indicating its role as an assistive tool rather than a fully autonomous system.
Example Output
The model takes raw receipt text and produces a detailed JSON output, including supplier, invoice_no, date, items (with product_raw, qty_cases, qty_units, unit_cost, total), subtotal, discount, gst, and net_total.