---
license: apache-2.0
viewer: false
---

# Dowload
```bash
wget "https://huggingface.co/datasets/Sterzhang/P-Bench-Choice/resolve/main/p-bench-qa.jsonl?download=true" -O p-bench-qa.jsonl
wget "https://huggingface.co/datasets/Sterzhang/P-Bench-Choice/resolve/main/P-Bench-imgs.tar.gz?download=true" -O P-Bench-imgs.tar.gz
mkdir P-Bench-imgs
tar -xzvf P-Bench-imgs.tar.gz -C P-Bench-imgs
```
## Image Folder (P-Bench-imgs)
```
P-Bench-imgs/
├── aug-face/
│   └── coco/
├── coco/
├── concat-person/
│   ├── 2person/
│   └── 3person/
├── crop-face/
│   ├── coco/
│   └── flickr30k/
└── flickr30k/
```

## JSONL file (p-bench-qa.jsonl)
```json
{"image": ["crop-face/coco/000000394849/3.jpg", "coco/000000394849.jpg"], "question": "<image>\nThis photo shows Sonia.\n<image>\nAnswer the following question based on this image:\nWhat is the color of Sonia's hair?\nOptions: A. blonde.; B. dark.; C. red.; D. pink.; E. I don't know Sonia.; F. Sonia is not in this image.", "answer": "B. dark.", "type": "normal", "cnt": "5"}
```
### Field Descriptions

- **`image`**: A list of image file names used in the sample. Each `<image>` token in the question field corresponds to an image in this list in order.  
  Each element in the list is a **relative path with respect to the `P-Bench-imgs/` directory**.

- **`question`**: A multi-modal question string that includes `<image>` placeholders and natural language text.

- **`answer`**: The ground-truth answer for this question. It must match one of the options listed in the question, and follows the format `"A. xxx."`.

- **`type`**: Indicates the task type, such as `"2person"`.

- **`cnt`**: Specifies the number of people present in the image referenced by the scene image (the **last element** of the `image` list).