ligaments-dev/Qwen2.5-1.5B-Instruct-itr-finetuned

Fine-tuned version of Qwen/Qwen2.5-1.5B-Instruct for Indian Income Tax Return (ITR) structured JSON extraction. The LoRA adapter has been merged into the base model weights (fused model).

Model Details

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Fine-tuning method: LoRA (rank=16, scale=32, dropout=0.05)
• Framework: MLX-LM v0.31.3 (Apple Silicon)
• Task: Extract structured JSON from ITR documents (ITR-1, ITR-2, ITR-3, ITR-4)
• Training: 3 epochs, 1500 iterations, lr=2e-5 (cosine decay), batch size=1 with grad accumulation=4
• Developed by: Ligaments AI

Evaluation Results

Evaluated on 49 held-out ITR examples:

Metric	Pass Rate
JSON Validity	98.0%
Form Type Match	98.0%
Numeric Sums Correct	98.0%
Boolean Y/N Only	98.0%
Date YYYY-MM-DD Format	98.0%
State/Country Numeric Codes	98.0%
No Round Numbers	81.6%

Usage

pip install mlx-lm

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("ligaments-dev/Qwen2.5-1.5B-Instruct-itr-finetuned")
sampler = make_sampler(temp=0.1)

messages = [
    {"role": "system", "content": "You are an ITR JSON extraction assistant..."},
    {"role": "user", "content": "<your ITR document text here>"}
]

prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_dict=False
)

response = generate(model, tokenizer, prompt=prompt, sampler=sampler, max_tokens=4096, verbose=True)

Intended Use

• Extracting structured financial data from Indian ITR documents for MSME lending workflows
• Automating credit risk assessment pipelines
• Not intended for general-purpose tax advice or legal decisions