2026-07-08 08:57:50 +00:00

148 lines
3.4 KiB
Markdown

# POPE
## Overview
POPE (Polling-based Object Probing Evaluation) is a benchmark specifically designed to evaluate object hallucination in Large Vision-Language Models (LVLMs). It tests models' ability to accurately identify objects present in images through yes/no questions.
## Task Description
- **Task Type**: Object Hallucination Detection (Yes/No Q&A)
- **Input**: Image with question "Is there a [object] in the image?"
- **Output**: YES or NO answer
- **Focus**: Measuring accuracy vs. hallucination rate
## Key Features
- Three sampling strategies: random, popular, adversarial
- Tests for false positive object claims (hallucination)
- Based on MSCOCO images
- Simple yes/no question format for objective evaluation
- Measures alignment between model responses and visual content
## Evaluation Notes
- Default configuration uses **0-shot** evaluation
- Five metrics: accuracy, precision, recall, F1 score, yes_ratio
- F1 score is the primary aggregation metric
- Three subsets: `popular`, `adversarial`, `random`
- "Popular" and "adversarial" subsets are more challenging
- yes_ratio indicates model's tendency to answer "yes"
## Properties
| Property | Value |
|----------|-------|
| **Benchmark Name** | `pope` |
| **Dataset ID** | [lmms-lab/POPE](https://modelscope.cn/datasets/lmms-lab/POPE/summary) |
| **Paper** | N/A |
| **Tags** | `Hallucination`, `MultiModal`, `Yes/No` |
| **Metrics** | `accuracy`, `precision`, `recall`, `f1_score`, `yes_ratio` |
| **Default Shots** | 0-shot |
| **Evaluation Split** | `N/A` |
| **Aggregation** | `f1` |
## Data Statistics
| Metric | Value |
|--------|-------|
| Total Samples | 9,000 |
| Prompt Length (Mean) | 79.4 chars |
| Prompt Length (Min/Max) | 75 / 87 chars |
**Per-Subset Statistics:**
| Subset | Samples | Prompt Mean | Prompt Min | Prompt Max |
|--------|---------|-------------|------------|------------|
| `popular` | 3,000 | 79.27 | 75 | 87 |
| `adversarial` | 3,000 | 79.36 | 75 | 87 |
| `random` | 3,000 | 79.59 | 75 | 87 |
**Image Statistics:**
| Metric | Value |
|--------|-------|
| Total Images | 9,000 |
| Images per Sample | min: 1, max: 1, mean: 1 |
| Resolution Range | 500x243 - 640x640 |
| Formats | jpeg |
## Sample Example
**Subset**: `popular`
```json
{
"input": [
{
"id": "8847a5a3",
"content": [
{
"text": "Is there a snowboard in the image?\nPlease answer YES or NO without an explanation."
},
{
"image": "[BASE64_IMAGE: png, ~87.2KB]"
}
]
}
],
"target": "YES",
"id": 0,
"group_id": 0,
"metadata": {
"id": "3000",
"answer": "YES",
"category": "popular",
"question_id": "1"
}
}
```
## 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 pope \
--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=['pope'],
dataset_args={
'pope': {
# subset_list: ['popular', 'adversarial', 'random'] # optional, evaluate specific subsets
}
},
limit=10, # Remove this line for formal evaluation
)
run_task(task_cfg=task_cfg)
```