---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 103857954.75
    num_examples: 3498
  - name: validation
    num_bytes: 21699876.0
    num_examples: 740
  - name: test
    num_bytes: 23003112.0
    num_examples: 760
  download_size: 147406131
  dataset_size: 148560942.75
---

# Qwen2.5-VL Product Classification Dataset

This dataset is designed for fine-tuning the Qwen2.5-VL model on product classification tasks.

## Dataset Description

The dataset consists of product images with 10 balanced categories:
- 蛋白粉 (Protein powder)
- 纸巾 (Tissue paper)
- 衣服 (Clothes)
- 膏药贴 (Medicinal plasters)
- 胶囊 (Capsules)
- 电脑 (Computer)
- 手表 (Watch)
- 弓箭 (Bow and arrow)
- 医疗器械 (Medical equipment)
- 包 (Bag)

Each item has two types of samples:
1. Image classification: Given a product image, determine its category
2. Title classification: Given a product title, determine its category

## Dataset Structure

This dataset follows the ShareGPT format:

For image classification items:
```
{
    'id': 'classification_[unique_id]',
    'image_path': PIL.Image.Image(...),  # The product image
    'item_type': 'image_classification',
    'conversations': [
        {'from': 'human', 'value': 'Prompt text with <image> token'},
        {'from': 'assistant', 'value': 'Model expected response (class label)'}
    ]
}
```

For title classification items:
```
{
    'id': 'title_classification_[unique_id]',
    'item_type': 'title_classification',  
    'conversations': [
        {'from': 'human', 'value': 'Prompt text with product title'},
        {'from': 'assistant', 'value': 'Model expected response (class label)'}
    ]
}
```

## Dataset Statistics

- Train samples: 3498
- Validation samples: 740
- Test samples: 760

## Usage

```python
from datasets import load_dataset

# Load dataset from Hugging Face Hub
dataset = load_dataset("BrightXiaoHan/tari-product-image")

# Access an example
example = dataset["train"][0]
print(example["conversations"][0]['value'])
print(example["conversations"][1]['value'])
if "image_path" in example:
    image = example["image_path"]
```

This dataset is intended for fine-tuning Qwen2.5-VL models for product classification tasks.