README.md

---
dataset_info:
- config_name: default
  features:
  - name: image
    dtype: image
  - name: category
    dtype: string
  - name: question_id
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: int64
  splits:
  - name: test
    num_bytes: 281192667
    num_examples: 191
  download_size: 281175185
  dataset_size: 281192667
- config_name: direct_attributes
  features:
  - name: image
    dtype: image
  - name: category
    dtype: string
  - name: question_id
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: int64
  splits:
  - name: test
    num_bytes: 169304473
    num_examples: 115
  download_size: 168837509
  dataset_size: 169304473
- config_name: relative_position
  features:
  - name: image
    dtype: image
  - name: category
    dtype: string
  - name: question_id
    dtype: string
  - name: question
    dtype: string
  - name: choices
    list: string
  - name: answer
    dtype: int64
  splits:
  - name: test
    num_bytes: 111888168
    num_examples: 76
  download_size: 112339908
  dataset_size: 111888168
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
- config_name: direct_attributes
  data_files:
  - split: test
    path: direct_attributes/test-*
- config_name: relative_position
  data_files:
  - split: test
    path: relative_position/test-*
---
# vstar_bench (Reformatted)

This dataset is a **reformatted version** of
[`craigwu/vstar_bench`](https://huggingface.co/datasets/craigwu/vstar_bench).

The underlying data (images, questions, choices, and answers) is **unchanged**.
Only the representation has been normalized to make it easier to use in
standard multiple-choice (MCQ) and instruction-following evaluation pipelines.

---

## Dataset Format

Each example contains:

- `question` (`string`): question text  
- `choices` (`List[str]`): answer options (no letter prefixes)  
- `answer` (`int`): correct option index (0-based)  
- `image` (`Image`): associated image  

---

## Reconstructing MCQ Prompt

The original letter-based MCQ format can be reconstructed as follows:

```python
from datasets import load_dataset

ds = load_dataset("ohjoonhee/vstar_bench")
row = ds["test"][0]

question = row["question"]
choices = row["choices"]
answer = row["answer"]

post_prompt = "Answer with the option's letter from the given choices directly."

choices = [f"({chr(i + ord('A'))}) {choice}" for i, choice in enumerate(choices)]
text = "\n".join([question] + choices + [post_prompt])
print(text)

label = ["A", "B", "C", "D"][answer]
print(label)
```

## Notes
- Fully reversible to the original dataset format
- No samples were modified, added, or removed

For the original dataset, see:
[`craigwu/vstar_bench`](https://huggingface.co/datasets/craigwu/vstar_bench/viewer/default/test)