LoRA Fine-Tuned Qwen2.5-1.5B for Fraud Risk Classification

This model is a LoRA fine-tuned version of Qwen2.5-1.5B-Instruct, adapted to classify KYC/transaction notes as suspicious or not_suspicious, with a one-line reason — inspired by real-world fraud detection and identity verification use cases.

Model Details

Base model: Qwen/Qwen2.5-1.5B-Instruct
Fine-tuning method: LoRA (PEFT), applied to attention layers (q_proj, v_proj), r=16, alpha=32
Quantization: Loaded in 4-bit (NF4) during training
Training data: ~297 synthetically generated labeled examples covering identity mismatches, unusual transfer patterns, dormant-account activity, and document tampering, alongside routine legitimate transactions
Trained on: Free Google Colab T4 GPU, 3 epochs (~96 seconds total)

Results

Stage	Accuracy
Zero-shot baseline (no fine-tuning)	~60-70% (biased toward false positives)
After LoRA fine-tuning	100% (53/53 on held-out test set)

The base model consistently over-flagged routine transactions (groceries, rent, tax refunds) as suspicious. Fine-tuning on balanced examples corrected this bias.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base_model, "ishank9/lora-fraud-classifier")
tokenizer = AutoTokenizer.from_pretrained("ishank9/lora-fraud-classifier")

prompt = "Classify the following note as 'suspicious' or 'not_suspicious' and give a one-line reason.\n\nNote: A dormant account suddenly received ₹900,000 and transferred it out the same day.\n\nAnswer:"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Limitations

This model was trained on a relatively small, LLM-generated synthetic dataset. While it performs strongly on similarly-styled held-out examples, this reflects pattern-learning on clean synthetic data rather than a guarantee of real-world generalization to messier, real transaction data.

Model tree for ishank9/lora-fraud-classifier

Base model

Qwen/Qwen2.5-1.5B

Finetuned

Qwen/Qwen2.5-1.5B-Instruct

Adapter

(1261)

this model