2026-07-08 08:57:50 +00:00

149 lines
6.3 KiB
Markdown

# Multi-IF
## Overview
Multi-IF is a benchmark designed to evaluate LLM capabilities in multi-turn instruction following within a multilingual environment. It tests the ability to follow complex instructions across multiple conversation turns in different languages.
## Task Description
- **Task Type**: Multi-Turn Multilingual Instruction Following
- **Input**: Multi-turn conversation with instructions
- **Output**: Responses following given instructions
- **Domains**: Instruction following, multilingual understanding
## Key Features
- 11 supported languages: Chinese, English, German, Italian, Vietnamese, Spanish, Hindi, Portuguese, French, Thai, Russian
- Multi-turn conversation evaluation (up to 3 turns)
- Tests instruction following in multilingual contexts
- Both strict and loose evaluation metrics
- Prompt-level and instruction-level scoring
## Evaluation Notes
- Default evaluation uses the **train** split
- Configurable max_turns (1-3, default: 3)
- Four metrics tracked:
- `prompt_level_strict/loose`: Strict/loose prompt-level accuracy
- `inst_level_strict/loose`: Strict/loose instruction-level accuracy
- Requires: nltk, langdetect, emoji (for Chinese), pythainlp (for Thai)
## Properties
| Property | Value |
|----------|-------|
| **Benchmark Name** | `multi_if` |
| **Dataset ID** | [facebook/Multi-IF](https://modelscope.cn/datasets/facebook/Multi-IF/summary) |
| **Paper** | N/A |
| **Tags** | `InstructionFollowing`, `MultiLingual`, `MultiTurn` |
| **Metrics** | `prompt_level_strict`, `inst_level_strict`, `prompt_level_loose`, `inst_level_loose` |
| **Default Shots** | 0-shot |
| **Evaluation Split** | `train` |
## Data Statistics
| Metric | Value |
|--------|-------|
| Total Samples | 4,501 |
| Prompt Length (Mean) | 0 chars |
| Prompt Length (Min/Max) | 0 / 0 chars |
**Per-Subset Statistics:**
| Subset | Samples | Prompt Mean | Prompt Min | Prompt Max |
|--------|---------|-------------|------------|------------|
| `Chinese` | 454 | 0 | 0 | 0 |
| `English` | 909 | 0 | 0 | 0 |
| `Italian` | 493 | 0 | 0 | 0 |
| `Spanish` | 516 | 0 | 0 | 0 |
| `Hindi` | 542 | 0 | 0 | 0 |
| `Portuguese` | 524 | 0 | 0 | 0 |
| `French` | 548 | 0 | 0 | 0 |
| `Russian` | 515 | 0 | 0 | 0 |
## Sample Example
**Subset**: `Chinese`
```json
{
"input": [
{
"id": "cb0c68fe",
"content": ""
}
],
"target": "",
"id": 0,
"group_id": 0,
"subset_key": "Chinese",
"metadata": {
"turns": null,
"responses": null,
"turn_1_prompt": "{\"role\": \"user\", \"content\": \"\\u5199\\u4e00\\u4e2a300+\\u5b57\\u7684\\u603b\\u7ed3\\u5173\\u4e8e\\u7ef4\\u57fa\\u767e\\u79d1\\u9875\\u9762\\\"https://en.wikipedia.org/wiki/Raymond_III,_Count_of_Tripoli\\\"\\uff0c\\u4e0d\\u8981\\u4f7f\\u7528\\u4efb\\u4f55\\u9017\\u53f7\\uff0c\\u5e76\\u4e14\\u81f3\\u5c11\\u7528markdown\\u683c\\u5f0f\\u7a81\\u51fa3\\u4e2a\\u6709\\u6807\\u9898\\u7684\\u90e8\\u5206\\uff0c\\u4f8b\\u5982*\\u7a81\\u51fa\\u90e8\\u52061*\\uff0c*\\u7a81\\u51fa\\u90e8\\u52062*\\uff0c*\\u7a81\\u51fa\\u90e8\\u52063*\\u3002\"}",
"turn_1_instruction_id_list": "[\"punctuation:no_comma\", \"detectable_format:number_highlighted_sections\", \"length_constraints:number_words\"]",
"turn_1_kwargs": "[\"{}\", \"{\\\"num_highlights\\\": 3}\", \"{\\\"relation\\\": \\\"at least\\\", \\\"num_words\\\": 300}\"]",
"turn_2_prompt": "{\"role\": \"user\", \"content\": \"\\u4f60\\u7684\\u56de\\u7b54\\u5e94\\u8be5\\u5305\\u542b\\u4ee5\\u4e0b\\u5173\\u952e\\u8bcd\\uff1a\\u5341\\u5b57\\u519b\\uff0c\\u9ece\\u5df4\\u5ae9\\uff0c\\u7a46\\u65af\\u6797\\u3002\"}",
"turn_2_instruction_id_list": "[\"punctuation:no_comma\", \"detectable_format:number_highlighted_sections\", \"length_constraints:number_words\", \"keywords:existence\"]",
"turn_2_kwargs": "[\"{}\", \"{\\\"num_highlights\\\": 3}\", \"{\\\"relation\\\": \\\"at least\\\", \\\"num_words\\\": 300}\", \"{\\\"keywords\\\": [\\\"\\\\u5341\\\\u5b57\\\\u519b\\\", \\\"\\\\u9ece\\\\u5df4\\\\u5ae9\\\", \\\"\\\\u7a46\\\\u65af\\\\u6797\\\"]}\"]",
"turn_3_prompt": "{\"role\": \"user\", \"content\": \"\\u4f60\\u7684\\u56de\\u7b54\\u5e94\\u8be5\\u4ee5\\u201c\\u8fd9\\u4e2a\\u6982\\u8ff0\\u63d0\\u4f9b\\u4e86\\u5341\\u5b57\\u519b\\u65f6\\u4ee3\\u653f\\u6cbb\\u548c\\u6218\\u4e89\\u7684\\u590d\\u6742\\u6027\\u548c\\u9634\\u8c0b\\u7684\\u6982\\u89c8\\u3002\\u201d\\u8fd9\\u4e2a\\u786e\\u5207\\u7684\\u77ed\\u8bed\\u7ed3\\u5c3e\\uff0c\\u4e0d\\u5141\\u8bb8\\u6709\\u5176\\u4ed6\\u6587\\u5b57\\u51fa\\u73b0\\u5728\\u8fd9\\u4e2a\\u77ed\\u8bed\\u540e\\u9762\\u3002\"}",
"turn_3_instruction_id_list": "[\"punctuation:no_comma\", \"detectable_format:number_highlighted_sections\", \"length_constraints:number_words\", \"keywords:existence\", \"startend:end_checker\"]",
"turn_3_kwargs": "[\"{}\", \"{\\\"num_highlights\\\": 3}\", \"{\\\"relation\\\": \\\"at least\\\", \\\"num_words\\\": 300}\", \"{\\\"keywords\\\": [\\\"\\\\u5341\\\\u5b57\\\\u519b\\\", \\\"\\\\u9ece\\\\u5df4\\\\u5ae9\\\", \\\"\\\\u7a46\\\\u65af\\\\u6797\\\"]}\", \"{\\\"end_phrase\\\": \\\"\\\\u8fd9\\\\u4e2a\\\\u6982\\\\u8ff0\\\\u63d0\\\\u4f9b\\\\u4e86\\\\u5341\\\\u5b57\\\\u519b\\\\u65f6\\\\u4ee3\\\\u653f\\\\u6cbb\\\\u548c\\\\u6218\\\\u4e89\\\\u7684\\\\u590d\\\\u6742\\\\u6027\\\\u548c\\\\u9634\\\\u8c0b\\\\u7684\\\\u6982\\\\u89c8\\\\u3002\\\"}\"]",
"key": "1000:1:zh",
"turn_index": 0,
"language": "Chinese"
}
}
```
## Prompt Template
*No prompt template defined.*
## Extra Parameters
| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `max_turns` | `int` | `3` | Maximum number of interactive turns to evaluate (1-3). Choices: [1, 2, 3] |
## Usage
### Using CLI
```bash
evalscope eval \
--model YOUR_MODEL \
--api-url OPENAI_API_COMPAT_URL \
--api-key EMPTY_TOKEN \
--datasets multi_if \
--limit 10 # Remove this line for formal evaluation
```
### Using 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=['multi_if'],
dataset_args={
'multi_if': {
# subset_list: ['Chinese', 'English', 'Italian'] # optional, evaluate specific subsets
# extra_params: {} # uses default extra parameters
}
},
limit=10, # Remove this line for formal evaluation
)
run_task(task_cfg=task_cfg)
```