117 lines
3.0 KiB
Markdown
117 lines
3.0 KiB
Markdown
# QASC
|
||
|
||
## 概述
|
||
|
||
QASC（Question Answering via Sentence Composition，基于句子组合的问题回答）是一个专注于多跳句子组合的问题回答数据集。它包含 9,980 个面向小学科学的 8 选 1 多项选择题，要求模型组合多个事实才能得出正确答案。
|
||
|
||
## 任务描述
|
||
|
||
- **任务类型**：多跳科学问答（多项选择）
|
||
- **输入**：包含 8 个选项的科学问题
|
||
- **输出**：正确答案对应的字母
|
||
- **重点**：句子组合与多跳推理
|
||
|
||
## 主要特点
|
||
|
||
- 包含 9,980 个小学科学问题
|
||
- 采用 8 选 1 的多项选择格式
|
||
- 需要组合两个事实才能作答
|
||
- 测试模型在科学知识上的多跳推理能力
|
||
- 每个问题均标注了支持性事实
|
||
|
||
## 评估说明
|
||
|
||
- 默认配置使用 **0-shot** 评估
|
||
- 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
|
||
- 适用于评估组合推理能力
|
||
|
||
## 属性
|
||
|
||
| 属性 | 值 |
|
||
|----------|-------|
|
||
| **基准测试名称** | `qasc` |
|
||
| **数据集ID** | [extraordinarylab/qasc](https://modelscope.cn/datasets/extraordinarylab/qasc/summary) |
|
||
| **论文** | N/A |
|
||
| **标签** | `Knowledge`, `MCQ` |
|
||
| **指标** | `acc` |
|
||
| **默认示例数量** | 0-shot |
|
||
| **评估划分** | `validation` |
|
||
|
||
## 数据统计
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总样本数 | 926 |
|
||
| 提示词长度（平均） | 362.58 字符 |
|
||
| 提示词长度（最小/最大） | 283 / 505 字符 |
|
||
|
||
## 样例示例
|
||
|
||
**子集**: `default`
|
||
|
||
```json
|
||
{
|
||
"input": [
|
||
{
|
||
"id": "d92a545e",
|
||
"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,F,G,H.\n\nClimate is generally described in terms of what?\n\nA) sand\nB) occurs over a wide range\nC) forests\nD) Global warming\nE) rapid changes occur\nF) local weather conditions\nG) measure of motion\nH) city life"
|
||
}
|
||
],
|
||
"choices": [
|
||
"sand",
|
||
"occurs over a wide range",
|
||
"forests",
|
||
"Global warming",
|
||
"rapid changes occur",
|
||
"local weather conditions",
|
||
"measure of motion",
|
||
"city life"
|
||
],
|
||
"target": "F",
|
||
"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
|
||
|
||
### 使用命令行（CLI）
|
||
|
||
```bash
|
||
evalscope eval \
|
||
--model YOUR_MODEL \
|
||
--api-url OPENAI_API_COMPAT_URL \
|
||
--api-key EMPTY_TOKEN \
|
||
--datasets qasc \
|
||
--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=['qasc'],
|
||
limit=10, # 正式评估时请删除此行
|
||
)
|
||
|
||
run_task(task_cfg=task_cfg)
|
||
``` |