150 lines
4.2 KiB
Markdown
150 lines
4.2 KiB
Markdown
# MGSM
|
||
|
||
## 概述
|
||
|
||
MGSM（Multilingual Grade School Math，多语言小学数学）是一个用于评估语言模型多语言数学推理能力的基准测试。它将 GSM8K 扩展至 11 种类型学上多样化的语言，以检验模型是否能在不同语言中执行思维链（chain-of-thought）推理。
|
||
|
||
## 任务描述
|
||
|
||
- **任务类型**：多语言数学应用题求解
|
||
- **输入**：11 种语言之一的小学数学应用题
|
||
- **输出**：包含逐步推理过程和数值答案的解答
|
||
- **语言**：英语、西班牙语、法语、德语、俄语、中文、日语、泰语、斯瓦希里语、孟加拉语、泰卢固语
|
||
|
||
## 主要特点
|
||
|
||
- 每种语言包含 250 道题目（从 GSM8K 翻译而来）
|
||
- 覆盖 11 种类型学上多样化的语言，涵盖不同语系
|
||
- 测试模型的多语言思维链推理能力
|
||
- 各语言使用相同的问题内容，便于跨语言比较
|
||
- 旨在评估与语言无关的数学推理能力
|
||
|
||
## 评估说明
|
||
|
||
- 默认配置使用 **4-shot** 示例
|
||
- 答案应使用 `\boxed{}` 格式包裹，以便正确提取
|
||
- 可通过 `subset_list` 参数指定评估特定语言（例如 `['en', 'zh', 'ja']`）
|
||
- 支持跨语言性能比较
|
||
- 少样本（few-shot）示例从同语言的训练集（train split）中抽取
|
||
|
||
## 属性
|
||
|
||
| 属性 | 值 |
|
||
|----------|-------|
|
||
| **基准测试名称** | `mgsm` |
|
||
| **数据集ID** | [evalscope/mgsm](https://modelscope.cn/datasets/evalscope/mgsm/summary) |
|
||
| **论文** | N/A |
|
||
| **标签** | `Math`, `MultiLingual`, `Reasoning` |
|
||
| **指标** | `acc` |
|
||
| **默认少样本数量** | 4-shot |
|
||
| **评估集** | `test` |
|
||
| **训练集** | `train` |
|
||
|
||
## 数据统计
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总样本数 | 2,750 |
|
||
| 提示词长度（平均） | 1742.98 字符 |
|
||
| 提示词长度（最小/最大） | 791 / 2464 字符 |
|
||
|
||
**各子集统计数据：**
|
||
|
||
| 子集 | 样本数 | 提示平均长度 | 提示最小长度 | 提示最大长度 |
|
||
|--------|---------|-------------|------------|------------|
|
||
| `en` | 250 | 1790.71 | 1637 | 2165 |
|
||
| `es` | 250 | 1940.02 | 1773 | 2371 |
|
||
| `fr` | 250 | 2047 | 1878 | 2440 |
|
||
| `de` | 250 | 1963.9 | 1792 | 2386 |
|
||
| `ru` | 250 | 1831.66 | 1667 | 2214 |
|
||
| `zh` | 250 | 842.16 | 791 | 946 |
|
||
| `ja` | 250 | 1102.33 | 1035 | 1248 |
|
||
| `th` | 250 | 1835.53 | 1699 | 2135 |
|
||
| `sw` | 250 | 1953.48 | 1780 | 2354 |
|
||
| `bn` | 250 | 1759.28 | 1601 | 2106 |
|
||
| `te` | 250 | 2106.77 | 1939 | 2464 |
|
||
|
||
## 样例示例
|
||
|
||
**子集**: `en`
|
||
|
||
```json
|
||
{
|
||
"input": [
|
||
{
|
||
"id": "d67cb3cf",
|
||
"content": "Here are some examples of how to solve similar problems:\n\nQuestion: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?\n\nReasoning:\nStep-by-Step Answer: Roger sta ... [TRUNCATED] ... every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market?\nPlease reason step by step, and put your final answer within \\boxed{}.\n\n"
|
||
}
|
||
],
|
||
"target": "18",
|
||
"id": 0,
|
||
"group_id": 0,
|
||
"metadata": {
|
||
"reasoning": null,
|
||
"equation_solution": null
|
||
}
|
||
}
|
||
```
|
||
|
||
*注：部分内容为显示目的已截断。*
|
||
|
||
## 提示模板
|
||
|
||
**提示模板：**
|
||
```text
|
||
{question}
|
||
Please reason step by step, and put your final answer within \boxed{}.
|
||
|
||
|
||
```
|
||
|
||
<details>
|
||
<summary>少样本模板</summary>
|
||
|
||
```text
|
||
Here are some examples of how to solve similar problems:
|
||
|
||
{fewshot}
|
||
|
||
{question}
|
||
Please reason step by step, and put your final answer within \boxed{}.
|
||
|
||
|
||
```
|
||
|
||
</details>
|
||
|
||
## 使用方法
|
||
|
||
### 使用 CLI
|
||
|
||
```bash
|
||
evalscope eval \
|
||
--model YOUR_MODEL \
|
||
--api-url OPENAI_API_COMPAT_URL \
|
||
--api-key EMPTY_TOKEN \
|
||
--datasets mgsm \
|
||
--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=['mgsm'],
|
||
dataset_args={
|
||
'mgsm': {
|
||
# subset_list: ['en', 'es', 'fr'] # 可选，用于评估特定子集
|
||
}
|
||
},
|
||
limit=10, # 正式评估时请删除此行
|
||
)
|
||
|
||
run_task(task_cfg=task_cfg)
|
||
``` |