Qwen3 Indian Address Parser

About

This LoRA adapter, fine-tuned on Qwen3-0.6B, specialises in parsing unstructured Indian address strings into structured JSON with 13 distinct fields: house number, building, street, locality, sub-district, district, city, state, pincode, and more. Trained on 5,008 gold-labeled records sourced from Indian MCA corporate data and financial institution branches, it achieves 82.4% per-field accuracy with pincode, state, and district extraction consistently exceeding 90%. Available as a Python package via pip, with dual loading paths for PEFT (cross-platform) and MLX (Apple Silicon). Licensed Apache 2.0.

Features

• Parses raw Indian addresses into 13 structured JSON fields
• 82.4% average per-field accuracy; 90%+ on pincode, state & district
• 100% JSON parse rate across evaluation samples
• Fine-tuned on 5,008 gold-labeled Indian address records
• PEFT (cross-platform) and MLX (Apple Silicon) loading paths
• pip-installable Python package for easy integration
• Apache 2.0 licence — free for commercial use

Tags