evalstone/evalscope/docs/en/benchmarks/general_vqa.md

General-VQA

Overview

General-VQA is a customizable visual question answering benchmark for evaluating multimodal models. It supports OpenAI-compatible message format with flexible image/video/audio input (local paths, URLs, or base64).

Task Description

• Task Type: Visual Question Answering
• Input: Images/videos/audio + questions in OpenAI chat format
• Output: Free-form text answer
• Flexibility: Supports custom datasets via TSV/JSONL files

Key Features

• OpenAI-compatible message format
• Supports multiple image/video/audio input methods (path, URL, base64)
• Media placeholders: Use <image N> / <video N> / <audio N> in plain-text user messages with indexed media columns (image_1, video_1, audio_1, etc.) for a simpler data format
• Flexible evaluation with BLEU and Rouge metrics
• Custom dataset support via local file loading
• Extensible for various VQA use cases

Evaluation Notes

• Default configuration uses 0-shot evaluation
• Default metrics: BLEU, Rouge (Rouge-L-R as main score)
• Evaluates on test split
• See User Guide for dataset format

Properties

Property	Value
Benchmark Name	general_vqa
Dataset ID	general_vqa
Paper	N/A
Tags	Custom, MultiModal, QA
Metrics	BLEU, Rouge
Default Shots	0-shot
Evaluation Split	test

Data Statistics

Statistics not available.

Sample Example

Sample example not available.

Prompt Template

No prompt template defined.

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets general_vqa \
    --limit 10  # Remove this line for formal evaluation

Using 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=['general_vqa'],
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)