SatyamGhosh/qlora-consumer-complaint-extractor

Consumer Complaint Structured Extraction (QLoRA Adapter)

LoRA adapter fine-tuned on Qwen/Qwen2.5-3B-Instruct using 4-bit QLoRA.

The model converts consumer complaint narratives into structured JSON:

• product
• issue
• disputed_amount
• summary

Data

The dataset uses real complaints from the CFPB Consumer Complaint Database.

product and issue are the categories present in the CFPB complaint data.

disputed_amount and summary were extracted using deterministic regex/rule-based methods.

No LLM was used to construct the dataset.

Results

Held-out validation set: 60 examples

Metric	Baseline	Fine-tuned
JSON validity	100.0%	100.0%
Field accuracy	60.0%	69.2%

Usage

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

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-3B-Instruct",
    quantization_config=bnb_config,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base,
    "SatyamGhosh/qlora-consumer-complaint-extractor"
)

tokenizer = AutoTokenizer.from_pretrained(
    "SatyamGhosh/qlora-consumer-complaint-extractor"
)