---
base_model: unsloth/qwen3-4b-base-unsloth-bnb-4bit
library_name: peft
model_name: Qwen3-4B-AI-Review-Detector
tags:
- base_model:adapter:unsloth/qwen3-4b-base-unsloth-bnb-4bit
- lora
- sft
- transformers
- trl
- unsloth
- text-classification
- korean
licence: license
pipeline_tag: text-generation
license: apache-2.0
language:
- ko
---

# Qwen3-4B-AI-Review-Detector

This model is a fine-tuned version of [unsloth/qwen3-4b-base-unsloth-bnb-4bit](https://huggingface.co/unsloth/qwen3-4b-base-unsloth-bnb-4bit) designed to detect whether a Korean **cosmetics** review is **Human-Written (HWR)** or **LLM-Generated (LGR)**.

It has been trained using [TRL](https://github.com/huggingface/trl) and [Unsloth](https://github.com/unslothai/unsloth) for efficient fine-tuning.

## Model Details

*   **Base Model**: `unsloth/qwen3-4b-base-unsloth-bnb-4bit`
*   **Task**: Binary Classification (via Text Generation)
    *   **Class 0**: HWR (Human Written Review)
    *   **Class 1**: LGR (LLM Generated Review)
*   **Language**: Korean
*   **Domain**: Cosmetics / Beauty
*   **Training Method**: LoRA (Low-Rank Adaptation)

## Quick start

You can use the `pipeline` from the `transformers` library to run inference.

```python
from transformers import pipeline

# Load the model
# Replace 'None' with your Hugging Face model ID (e.g., "username/Qwen3-4B-AI-Review-Detector")
model_id = "jedimark/Qwen3-4B-AI-Review-Detector" 
generator = pipeline("text-generation", model=model_id, device_map="auto")

# Example review
review_text = "이 제품 정말 좋아요! 배송도 빠르고 품질도 만족합니다."

# Construct the prompt
prompt = f"""다음 리뷰 텍스트가 사람이 작성한 것인지(Human Written) LLM이 생성한 것인지 판단하여 분류하세요.
{review_text}

Classify this review into one of the following:
class 0: HWR (Human Written Review)
class 1: LGR (LLM Generated Review)

SOLUTION
The correct answer is: class"""

# Run inference
output = generator(prompt, max_new_tokens=1, return_full_text=False)[0]
print(f"Predicted Class: {output['generated_text']}")
# Output: 0 (Human Written) or 1 (LLM Generated)
```

## Training procedure

This model was trained using **SFT (Supervised Fine-Tuning)** with the following configuration:

*   **Dataset**: Custom dataset of Korean **cosmetics** reviews labeled as Human-Written (0) or LLM-Generated (1).
*   **Quantization**: 4-bit quantization using `bitsandbytes` (BnB) for memory efficiency.
*   **LoRA Configuration**:
    *   Rank (r): 16
    *   Alpha: 16
    *   Target Modules: `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` (All linear layers)
*   **Optimization**: Trained with `Unsloth` for faster training and lower memory usage.

### Framework versions

- Unsloth 2024.x
- PEFT 0.18.0
- TRL: 0.24.0
- Transformers: 4.57.2
- Pytorch: 2.9.0
- Datasets: 4.3.0
- Tokenizers: 0.22.1