137 lines
3.3 KiB
Markdown
137 lines
3.3 KiB
Markdown
# CCBench
|
||
|
||
|
||
## 概述
|
||
|
||
CCBench（Chinese Culture Bench）是 MMBench 的一个扩展，专门用于评估多模态模型对中华传统文化的理解能力。它通过视觉问答的形式，涵盖中华文化遗产的多个方面。
|
||
|
||
## 任务描述
|
||
|
||
- **任务类型**：视觉多项选择题问答（中华文化）
|
||
- **输入**：包含中华文化相关问题的图像
|
||
- **输出**：单个正确答案字母（A、B、C 或 D）
|
||
- **语言**：主要为中文内容
|
||
|
||
## 主要特点
|
||
|
||
- 聚焦中华传统文化相关问题
|
||
- 类别包括：书法、绘画、文物、饮食与服饰
|
||
- 历史人物、风景与建筑、草图推理、传统表演
|
||
- 测试结合视觉理解与文化知识的能力
|
||
- 基于 MMBench 评估框架的扩展
|
||
|
||
## 评估说明
|
||
|
||
- 默认配置使用 **0-shot** 评估
|
||
- 使用思维链（Chain-of-Thought, CoT）提示
|
||
- 在测试集（test split）上进行评估
|
||
- 使用简单准确率（accuracy）作为评分指标
|
||
- 要求模型同时具备视觉感知能力和文化知识
|
||
|
||
## 属性
|
||
|
||
| 属性 | 值 |
|
||
|----------|-------|
|
||
| **基准测试名称** | `cc_bench` |
|
||
| **数据集ID** | [lmms-lab/MMBench](https://modelscope.cn/datasets/lmms-lab/MMBench/summary) |
|
||
| **论文** | N/A |
|
||
| **标签** | `Knowledge`, `MCQ`, `MultiModal` |
|
||
| **指标** | `acc` |
|
||
| **默认示例数** | 0-shot |
|
||
| **评估划分** | `test` |
|
||
|
||
|
||
## 数据统计
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总样本数 | 2,040 |
|
||
| 提示词长度（平均） | 270.1 字符 |
|
||
| 提示词长度（最小/最大） | 254 / 394 字符 |
|
||
|
||
**图像统计信息：**
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总图像数 | 2,040 |
|
||
| 每样本图像数 | 最小: 1, 最大: 1, 平均: 1 |
|
||
| 分辨率范围 | 119x118 - 512x512 |
|
||
| 格式 | jpeg |
|
||
|
||
|
||
## 样例示例
|
||
|
||
**子集**: `cc`
|
||
|
||
```json
|
||
{
|
||
"input": [
|
||
{
|
||
"id": "2797f551",
|
||
"content": [
|
||
{
|
||
"text": "Answer the following multiple choice question. The last line of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of A,B,C,D. Think step by step before answering.\n\n图中所示建筑名称为？\n\nA) 天坛\nB) 故宫\nC) 黄鹤楼\nD) 少林寺"
|
||
},
|
||
{
|
||
"image": "[BASE64_IMAGE: jpeg, ~22.7KB]"
|
||
}
|
||
]
|
||
}
|
||
],
|
||
"choices": [
|
||
"天坛",
|
||
"故宫",
|
||
"黄鹤楼",
|
||
"少林寺"
|
||
],
|
||
"target": "A",
|
||
"id": 0,
|
||
"group_id": 0,
|
||
"metadata": {
|
||
"index": 0,
|
||
"category": "scenery_building",
|
||
"source": "https://zh.wikipedia.org/wiki/%E5%A4%A9%E5%9D%9B"
|
||
}
|
||
}
|
||
```
|
||
|
||
## 提示模板
|
||
|
||
**提示模板：**
|
||
```text
|
||
Answer the following multiple choice question. The last line of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of {letters}. Think step by step before answering.
|
||
|
||
{question}
|
||
|
||
{choices}
|
||
```
|
||
|
||
## 使用方法
|
||
|
||
### 使用 CLI
|
||
|
||
```bash
|
||
evalscope eval \
|
||
--model YOUR_MODEL \
|
||
--api-url OPENAI_API_COMPAT_URL \
|
||
--api-key EMPTY_TOKEN \
|
||
--datasets cc_bench \
|
||
--limit 10 # 正式评估时请删除此行
|
||
```
|
||
|
||
### 使用 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=['cc_bench'],
|
||
limit=10, # 正式评估时请删除此行
|
||
)
|
||
|
||
run_task(task_cfg=task_cfg)
|
||
``` |