README.md

---
pretty_name: WTB-Bench
language:
- en
license: apache-2.0
task_categories:
- visual-question-answering
task_ids:
- multiple-choice-qa
tags:
- benchmark
- visual-question-answering
- wind-turbine
- inspection
- defect-detection
- computer-vision
size_categories:
- 1K<n<10K
---

# WTB-Bench

WTB-Bench is a visual question answering benchmark for wind turbine inspection. Each example pairs a turbine image with a multiple-choice question, answer key, options, and task type.

## Dataset Structure

This repository uses the Hugging Face ImageFolder layout:

```text
test/
  metadata.jsonl
  images/
    <sha256>.jpg
```

`metadata.jsonl` contains one JSON object per question. The `file_name` field points to the relative image path.

## Fields

- `image`: The wind turbine inspection image loaded by Hugging Face Datasets from `file_name`.
- `question`: Multiple-choice question.
- `answer`: Correct option key, such as `A`, `B`, `C`, or `D`.
- `options`: Mapping from option key to option text.
- `task_type`: Question category, such as `existence`, `counting`, `classification`, `localization`, or `analysis`.

## Splits

- `test`: 1,200 QA entries over 288 unique images.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("withstaticTai/WTB-Bench")
example = dataset["test"][0]
```

## Intended Use

WTB-Bench is intended for evaluating visual question answering systems on wind turbine inspection scenarios, including defect existence, counting, classification, localization, and inspection implication analysis.

## Limitations

The dataset is provided as a benchmark test split. It should not be treated as a complete training corpus for wind turbine defect recognition. Model outputs should be interpreted with domain expertise before any operational inspection or maintenance decision.