gagan1985/qwen3-0.6b-indian-address-parser

Qwen3-0.6B LoRA — Indian Address Parser

A LoRA adapter fine-tuned on Qwen/Qwen3-0.6B that parses raw, unstructured Indian address strings into 13 structured fields, output as JSON.

Input:  "FLAT NO.32, UTTARA TOWERS, MG ROAD GUWAHATI , Kamrup Unclassified AS 781029"
Output: {"houseNumber": "FLAT NO.32", "houseName": "UTTARA TOWERS", "poi": null,
         "street": "MG ROAD", "subsubLocality": null, "subLocality": null, "locality": null,
         "village": null, "subDistrict": null, "district": "Kamrup", "city": "GUWAHATI",
         "state": "AS", "pincode": "781029"}

Also available as a Python package

pip install indian-address-parser

Source, CLI, and a benchmark comparison against shiprocket-ai/open-tinybert-indian-address-ner are on GitHub: innerkorehq/indian-address-parser.

Datasets

• gagan1985/indian-addresses-gold — the 4,834-example span-labeled training data behind this model
• gagan1985/indian-addresses-raw — the 4.37M-record raw, unlabeled corpus this gold set was drawn from

Two loading paths

Training was done with MLX (mlx-lm's lora command) on Apple Silicon. The repo root has a PEFT-format conversion of that adapter so it loads on any platform (CUDA, MPS, CPU) via standard transformers+peft — the mlx/ subfolder has the original MLX artifacts for Apple Silicon users. Both were verified to produce matching output (13/15 identical on a held-out spot check; the 2 differences landed on fields already noted as ambiguous below — consistent with floating-point differences between backends on a near-tied decision, not a conversion error).

Option A — PEFT (transformers, any platform)

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "gagan1985/qwen3-0.6b-indian-address-parser"
tokenizer = AutoTokenizer.from_pretrained(repo)
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, repo)

SYSTEM_PROMPT = (
    "You are an Indian address parser. Given a raw address string, extract address "
    "fields and return them as a JSON object. Use null for fields not present in the "
    "address. Output only the JSON object, no explanation.\n\n"
    "Fields: houseNumber, houseName, poi, street, subsubLocality, subLocality, "
    "locality, village, subDistrict, district, city, state, pincode"
)
messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "Parse this address:\n<your address here>"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False, pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Or use the included inference.py / evaluate.py (download the repo, then python inference.py --model Qwen/Qwen3-0.6B --adapter . "<address>").

Option B — MLX (Apple Silicon)

mlx_lm.load's adapter_path only accepts a local directory, not an HF repo ID — so fetch the mlx/ subfolder first, then point at it locally:

from huggingface_hub import snapshot_download
import mlx_lm

local_dir = snapshot_download("gagan1985/qwen3-0.6b-indian-address-parser", allow_patterns=["mlx/*"])
model, tokenizer = mlx_lm.load("Qwen/Qwen3-0.6B", adapter_path=f"{local_dir}/mlx")

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "Parse this address:\n<your address here>"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
print(mlx_lm.generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=False))

Or use mlx/inference_mlx.py / mlx/evaluate_mlx.py from a local checkout.

Fields

houseNumber, houseName, poi, street, subsubLocality, subLocality,
locality, village, subDistrict, district, city, state, pincode

Training data

• 5,008 gold-labeled records: 9 human-reviewed (Label Studio) + 4,999 LLM-reviewed (deepseek/deepseek-v4-pro via OpenRouter, span-verified against the raw address text — any field value not found verbatim in the source string was dropped, not guessed)
• Sourced from two distinct raw formats: Indian MCA (Ministry of Corporate Affairs) company-registration addresses, and bank/business-correspondent branch addresses
• Split 4,266 train / 237 val / 237 test (deduplicated)

Training config

Parameter	Value
Base model	Qwen/Qwen3-0.6B (28 layers)
LoRA rank / alpha / dropout	16 / 32 / 0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Fine-tuned layers	16 (of 28)
Iterations	2000 (≈3.75 epochs at batch=8)
Learning rate	2e-4
Trainable params	5.77M / 596M (0.968%)

Evaluation (237 held-out test samples)

• JSON parse rate: 100%
• Overall exact match (all present fields): 19.0%
• Mean per-field accuracy: 82.4%

Field	Accuracy	Recall	Gold presence
pincode	100.0%	100.0%	100.0%
state	94.9%	96.2%	98.7%
district	94.9%	95.1%	68.8%
houseNumber	90.3%	84.5%	54.4%
city	90.3%	91.3%	62.9%
houseName	87.3%	88.5%	43.9%
subDistrict	86.1%	29.6%	11.4%
poi	84.0%	30.8%	16.5%
village	75.1%	43.8%	27.0%
subLocality	74.3%	21.6%	21.5%
subsubLocality	72.6%	31.2%	27.0%
street	64.6%	55.6%	53.2%
locality	56.5%	35.6%	43.9%

Known limitations

• locality/subLocality/subsubLocality/village overlap conceptually. These represent the same "named area, different granularity" concept, and disagreements are often the model extracting the correct substring into an adjacent bucket rather than a genuine extraction miss — the gold labels themselves are sometimes inconsistent here (the same place name occasionally appears in two of these fields simultaneously in gold).
• street sometimes over-absorbs multi-part location clusters (e.g. "KARBI PATH, ZOO NARENGI ROAD, BAMUNIMAIDAN" tagged entirely as street instead of being split across street/subsubLocality/subLocality), likely because that split pattern is underrepresented in training data (subsubLocality/subLocality gold presence is only ~21-27%).
• Source addresses occasionally contain data artifacts (e.g. duplicated substrings from the original MCA records, like "PEDANANDIPALLE AGRAHARAMPEDANANDIPALLE AGRAHARAM") that the model sometimes reproduces with minor copy errors.

Files in this repo

Root (PEFT format):

• adapter_model.safetensors + adapter_config.json — standard PEFT LoRA adapter
• tokenizer.json, tokenizer_config.json, vocab.json, merges.txt, special_tokens_map.json, added_tokens.json, chat_template.jinja — tokenizer files
• inference.py / evaluate.py — CLI scripts (single address, stdin/file batch, or full test-set evaluation) using transformers+peft
• config.py — shared constants (SYSTEM_PROMPT, field list, prompt template)

mlx/ (original MLX format, Apple Silicon):

• adapters.safetensors + adapter_config.json — mlx-lm's native adapter format (adapter_config.json here is mlx-lm's training-run metadata, not a PEFT LoraConfig)
• inference_mlx.py / evaluate_mlx.py — same CLI shape as the root scripts, via mlx-lm
• config.py — same constants, duplicated here so this subfolder is usable standalone