# VASEVQA Dataset

A Visual Question Answering dataset for ancient vase analysis and description.

## Dataset Overview

- **Total Images**: 31,773 vase images
- **Format**: Conversational QA pairs in LLaVA format
- **Domain**: Ancient Greek/Roman pottery and vases

## Data Structure

### Single-view Dataset
- **Training**: `data_train_single_llava_vasevl_v9.json` (9,354 samples)
- **Testing**: `data_test_single_llava_vasevl_v9.json` (2,339 samples)
- Format: One image per conversation

### Multi-view Dataset  
- **Training**: `data_train_multi_llava_vasevl_v6_fixed.json` (25,544 samples)
- **Testing**: `data_test_multi_llava_vasevl_v6_fixed.json` (6,229 samples)
- Format: Multiple images per conversation with proper `<image>` tokens

## Question Types

Each vase is annotated with questions about:
- Fabric (e.g., ATHENIAN)
- Technique (e.g., BLACK-FIGURE, RED-FIGURE)
- Shape name (e.g., LEKYTHOS, MASTOID)
- Provenance/origin
- Dating information
- Attribution details
- Decorative descriptions
- Overall comprehensive information

## JSON Format

```json
{
  "id": "sample_id",
  "images": "images/filename.jpg" | ["image1.jpg", "image2.jpg"],
  "conversations": [
    {
      "from": "human",
      "value": "<image>\nWhat is the fabric of the vase?"
    },
    {
      "from": "gpt", 
      "value": "The fabric of the vase is ATHENIAN."
    }
  ]
}
```

## Files

- `images/`: Directory containing 31,773 vase images
- `descriptions.csv`: Original metadata source
- `fix_multi_data.py`: Script for adding image tokens to multi-view data