140 lines
3.4 KiB
Markdown
140 lines
3.4 KiB
Markdown
# HallusionBench
|
|
|
|
|
|
## Overview
|
|
|
|
HallusionBench is an advanced diagnostic benchmark designed to evaluate image-context reasoning and detect hallucination tendencies in Large Vision-Language Models (LVLMs). It specifically tests models' susceptibility to language hallucination and visual illusion.
|
|
|
|
## Task Description
|
|
|
|
- **Task Type**: Hallucination Detection and Visual Reasoning
|
|
- **Input**: Image + yes/no question about image content
|
|
- **Output**: YES or NO answer
|
|
- **Domains**: Hallucination detection, visual reasoning, factual accuracy
|
|
|
|
## Key Features
|
|
|
|
- Specifically designed to probe hallucination behaviors
|
|
- Tests both language hallucination and visual illusion
|
|
- Organized by categories and subcategories for detailed analysis
|
|
- Uses grouped accuracy metrics for robust evaluation
|
|
- Questions require precise image-context reasoning
|
|
|
|
## Evaluation Notes
|
|
|
|
- Default evaluation uses the **image** split
|
|
- Multiple accuracy metrics:
|
|
- **aAcc**: Answer-level accuracy (per-question)
|
|
- **fAcc**: Figure-level accuracy (all questions per figure correct)
|
|
- **qAcc**: Question-level accuracy (grouped by question type)
|
|
- Requires simple YES/NO answers without explanation
|
|
- Aggregation at subcategory, category, and overall levels
|
|
|
|
|
|
## Properties
|
|
|
|
| Property | Value |
|
|
|----------|-------|
|
|
| **Benchmark Name** | `hallusion_bench` |
|
|
| **Dataset ID** | [lmms-lab/HallusionBench](https://modelscope.cn/datasets/lmms-lab/HallusionBench/summary) |
|
|
| **Paper** | N/A |
|
|
| **Tags** | `Hallucination`, `MultiModal`, `Yes/No` |
|
|
| **Metrics** | `aAcc`, `qAcc`, `fAcc` |
|
|
| **Default Shots** | 0-shot |
|
|
| **Evaluation Split** | `image` |
|
|
| **Aggregation** | `f1` |
|
|
|
|
|
|
## Data Statistics
|
|
|
|
| Metric | Value |
|
|
|--------|-------|
|
|
| Total Samples | 951 |
|
|
| Prompt Length (Mean) | 136.78 chars |
|
|
| Prompt Length (Min/Max) | 76 / 292 chars |
|
|
|
|
**Image Statistics:**
|
|
|
|
| Metric | Value |
|
|
|--------|-------|
|
|
| Total Images | 951 |
|
|
| Images per Sample | min: 1, max: 1, mean: 1 |
|
|
| Resolution Range | 388x56 - 5291x4536 |
|
|
| Formats | png |
|
|
|
|
|
|
## Sample Example
|
|
|
|
**Subset**: `default`
|
|
|
|
```json
|
|
{
|
|
"input": [
|
|
{
|
|
"id": "ba75d669",
|
|
"content": [
|
|
{
|
|
"text": "Is China, Hongkong SAR, the leading importing country of gold, silverware, and jewelry with the highest import value in 2018?\nPlease answer YES or NO without an explanation."
|
|
},
|
|
{
|
|
"image": "[BASE64_IMAGE: png, ~143.0KB]"
|
|
}
|
|
]
|
|
}
|
|
],
|
|
"target": "NO",
|
|
"id": 0,
|
|
"group_id": 0,
|
|
"metadata": {
|
|
"category": "VS",
|
|
"subcategory": "chart",
|
|
"visual_input": "1",
|
|
"set_id": "0",
|
|
"figure_id": "1",
|
|
"question_id": "0",
|
|
"gt_answer": "0",
|
|
"gt_answer_details": "Switzerland is the leading importing country of gold, silverware, and jewelry with the highest import value in 2018?"
|
|
}
|
|
}
|
|
```
|
|
|
|
## Prompt Template
|
|
|
|
**Prompt Template:**
|
|
```text
|
|
{question}
|
|
Please answer YES or NO without an explanation.
|
|
```
|
|
|
|
## Usage
|
|
|
|
### Using CLI
|
|
|
|
```bash
|
|
evalscope eval \
|
|
--model YOUR_MODEL \
|
|
--api-url OPENAI_API_COMPAT_URL \
|
|
--api-key EMPTY_TOKEN \
|
|
--datasets hallusion_bench \
|
|
--limit 10 # Remove this line for formal evaluation
|
|
```
|
|
|
|
### Using Python
|
|
|
|
```python
|
|
from evalscope import run_task
|
|
from evalscope.config import TaskConfig
|
|
|
|
task_cfg = TaskConfig(
|
|
model='YOUR_MODEL',
|
|
api_url='OPENAI_API_COMPAT_URL',
|
|
api_key='EMPTY_TOKEN',
|
|
datasets=['hallusion_bench'],
|
|
limit=10, # Remove this line for formal evaluation
|
|
)
|
|
|
|
run_task(task_cfg=task_cfg)
|
|
```
|
|
|
|
|