115 lines
2.9 KiB
Markdown
115 lines
2.9 KiB
Markdown
# CommonsenseQA
|
||
|
||
## 概述
|
||
|
||
CommonsenseQA 是一个用于评估 AI 模型回答需要常识推理问题能力的基准测试。这些问题旨在考察模型是否具备问题中未明确说明的背景常识知识。
|
||
|
||
## 任务描述
|
||
|
||
- **任务类型**：多项选择常识推理
|
||
- **输入**：包含 5 个选项、需要常识知识的问题
|
||
- **输出**：正确答案的字母（A-E）
|
||
- **重点**：世界知识与常识推理
|
||
|
||
## 主要特点
|
||
|
||
- 问题基于 ConceptNet 知识图谱生成
|
||
- 需要多种类型的常识知识
|
||
- 每个问题提供 5 个选项
|
||
- 测试对日常概念及其关系的推理能力
|
||
- 所有问题均经过人工验证
|
||
|
||
## 评估说明
|
||
|
||
- 默认配置使用 **0-shot** 评估
|
||
- 使用简单的多项选择提示方式
|
||
- 在验证集（validation split）上进行评估
|
||
- 使用简单准确率（accuracy）作为评估指标
|
||
|
||
## 属性
|
||
|
||
| 属性 | 值 |
|
||
|----------|-------|
|
||
| **基准测试名称** | `commonsense_qa` |
|
||
| **数据集 ID** | [extraordinarylab/commonsense-qa](https://modelscope.cn/datasets/extraordinarylab/commonsense-qa/summary) |
|
||
| **论文** | N/A |
|
||
| **标签** | `Commonsense`, `MCQ`, `Reasoning` |
|
||
| **指标** | `acc` |
|
||
| **默认示例数** | 0-shot |
|
||
| **评估划分** | `validation` |
|
||
|
||
|
||
## 数据统计
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总样本数 | 1,221 |
|
||
| 提示词长度（平均） | 326.11 字符 |
|
||
| 提示词长度（最小/最大） | 257 / 537 字符 |
|
||
|
||
## 样例示例
|
||
|
||
**子集**: `default`
|
||
|
||
```json
|
||
{
|
||
"input": [
|
||
{
|
||
"id": "c9b96cfc",
|
||
"content": "Answer the following multiple choice question. The entire content of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of A,B,C,D,E.\n\nA revolving door is convenient for two direction travel, but it also serves as a security measure at a what?\n\nA) bank\nB) library\nC) department store\nD) mall\nE) new york"
|
||
}
|
||
],
|
||
"choices": [
|
||
"bank",
|
||
"library",
|
||
"department store",
|
||
"mall",
|
||
"new york"
|
||
],
|
||
"target": "A",
|
||
"id": 0,
|
||
"group_id": 0,
|
||
"metadata": {}
|
||
}
|
||
```
|
||
|
||
## 提示模板
|
||
|
||
**提示模板：**
|
||
```text
|
||
Answer the following multiple choice question. The entire content of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of {letters}.
|
||
|
||
{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 commonsense_qa \
|
||
--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=['commonsense_qa'],
|
||
limit=10, # 正式评估时请删除此行
|
||
)
|
||
|
||
run_task(task_cfg=task_cfg)
|
||
``` |