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

139 lines
3.2 KiB
Markdown

# VQAv2
## Overview
VQAv2 is the balanced Visual Question Answering benchmark built on COCO images. It evaluates whether
multimodal models can answer open-ended natural-language questions grounded in image content.
## Task Description
- **Task Type**: Open-ended visual question answering
- **Input**: Image + natural-language question
- **Output**: Short answer phrase
- **Domains**: General image understanding, object recognition, counting, attributes, relations
## Evaluation Notes
- Default data source: `lmms-lab/VQAv2` on ModelScope, `validation` split
- Primary metric: **VQAv2 soft accuracy** over human annotator answers
- Also reports normalized exact match against the available answer set
- The adapter accepts common answer formats: list of strings, list of answer dicts, or `multiple_choice_answer`
## Properties
| Property | Value |
|----------|-------|
| **Benchmark Name** | `vqav2` |
| **Dataset ID** | [lmms-lab/VQAv2](https://modelscope.cn/datasets/lmms-lab/VQAv2/summary) |
| **Paper** | [Paper](https://arxiv.org/abs/1612.00837) |
| **Tags** | `MultiModal`, `QA` |
| **Metrics** | `vqa_score`, `exact_match` |
| **Default Shots** | 0-shot |
| **Evaluation Split** | `validation` |
## Data Statistics
| Metric | Value |
|--------|-------|
| Total Samples | 214,354 |
| Prompt Length (Mean) | 185.83 chars |
| Prompt Length (Min/Max) | 165 / 255 chars |
**Image Statistics:**
| Metric | Value |
|--------|-------|
| Total Images | 214,354 |
| Images per Sample | min: 1, max: 1, mean: 1 |
| Resolution Range | 120x120 - 640x640 |
| Formats | jpeg, png |
## Sample Example
**Subset**: `default`
```json
{
"input": [
{
"id": "7410f0b5",
"content": [
{
"text": "Answer the question according to the image using a short phrase.\nWhere is he looking?\nThe last line of your response should be of the form \"ANSWER: [ANSWER]\" (without quotes)."
},
{
"image": "[BASE64_IMAGE: jpeg, ~102.7KB]"
}
]
}
],
"target": "[\"down\", \"down\", \"at table\", \"skateboard\", \"down\", \"table\", \"down\", \"down\", \"down\", \"down\"]",
"id": 0,
"group_id": 0,
"metadata": {
"question": "Where is he looking?",
"answers": [
"down",
"down",
"at table",
"skateboard",
"down",
"table",
"down",
"down",
"down",
"down"
],
"multiple_choice_answer": "down",
"question_id": 262148000,
"question_type": "none of the above",
"answer_type": "other"
}
}
```
## Prompt Template
**Prompt Template:**
```text
Answer the question according to the image using a short phrase.
{question}
The last line of your response should be of the form "ANSWER: [ANSWER]" (without quotes).
```
## Usage
### Using CLI
```bash
evalscope eval \
--model YOUR_MODEL \
--api-url OPENAI_API_COMPAT_URL \
--api-key EMPTY_TOKEN \
--datasets vqav2 \
--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=['vqav2'],
limit=10, # Remove this line for formal evaluation
)
run_task(task_cfg=task_cfg)
```