IF-Verifier-Data

# Dataset Card for Dataset Name

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

- **Curated by:** Hao Peng@THUKEG
- **Language(s) (NLP):** English, Chinese
- **License:** apache-2.0

### Dataset Sources [optional]

<!-- Provide the basic links for the dataset. -->

- **Repository:** https://github.com/THU-KEG/VerIF
- **Paper:** https://arxiv.org/abs/2506.09942

## Uses

This data is used for training generative reward models for instruction-following.

## Dataset Structure
The data is in `jsonl` format, with each line being a json item with the following format:

```
{
"id": <data id>,
"messages": [
{"role": "user", "content": <user query>},
{"role": "assistant", "content": <response from QwQ 32B>}

]
}

```
## Dataset Creation

### Source Data

The original data is WildChat (https://huggingface.co/datasets/allenai/WildChat) and InfinityInstruct (https://huggingface.co/datasets/BAAI/Infinity-Instruct).

#### Data Collection and Processing

We first generate an additional **20,000** data instances as in [VerInstruct](https://huggingface.co/datasets/Wesleythu/Crab-VerIF). To ensure diversity, we additionally mined complex instructions from WildChat and Infinity Instruct~. Specifically, we use Qwen2.5-72B-Instruct to extract constraints from each instruction and classify them as hard or soft. For hard constraints, we adopt Qwen2.5-72B-Instruct to generate corresponding verification Python code scripts. For each instruction, we randomly sample a response from *6* different models, including Llama3.1-8B-Instruct, Llama-3.3-70B-Instruct, Qwen2.5-7B-Instruct, Qwen2.5-72B-Instruct, QwQ-32B, DeepSeek-R1-Distilled-Qwen-32B. We then adopt QwQ-32B to generate a step-by-step verification indicating whether the output satisfies the instruction for each instruction-response pair. As a result, we collect about $130$k instruction–response pairs with corresponding step-by-step verification.

For more details, please refer to our paper and out GitHub [repo](https://github.com/THU-KEG/VerIF).

## Citation

```
@misc{peng2025verif,
title={VerIF: Verification Engineering for Reinforcement Learning in Instruction Following},
author={Hao Peng and Yunjia Qi and Xiaozhi Wang and Bin Xu and Lei Hou and Juanzi Li},
year={2025},
eprint={2506.09942},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.09942},
}
```

## Dataset Card Contact

Please contact [[e-mail]] if you have any questions.

# 数据集卡片（Dataset Card）

## 数据集详情

### 数据集描述
<!-- 请提供该数据集的详细概述。 -->

- **数据整理者：** 郝鹏@清华知识工程组（THUKEG）
- **（自然语言处理领域）使用语言：** 英语、中文
- **许可证：** Apache-2.0

### 数据集来源（可选）
<!-- 请提供该数据集的基础链接。 -->

- **代码仓库：** https://github.com/THU-KEG/VerIF
- **相关论文：** https://arxiv.org/abs/2506.09942

## 数据集用途
本数据集用于训练面向指令遵循任务的生成式奖励模型。

## 数据集结构
本数据集采用JSON Lines（jsonl）格式存储，每一行均为一条符合以下格式的JSON条目：
json
{
"id": <数据编号>,
"messages": [
{"role": "user", "content": <用户查询内容>},
{"role": "assistant", "content": <QwQ 32B生成的助手回复>}

]
}

## 数据集构建

### 源数据
本数据集的原始数据为WildChat（https://huggingface.co/datasets/allenai/WildChat）与InfinityInstruct（https://huggingface.co/datasets/BAAI/Infinity-Instruct）。

#### 数据收集与处理流程
我们首先参考VerInstruct（https://huggingface.co/datasets/Wesleythu/Crab-VerIF）额外生成了20000条数据实例。为保证数据多样性，我们额外从WildChat与InfinityInstruct中挖掘复杂指令。具体而言，我们使用Qwen2.5-72B-Instruct从每条指令中提取约束条件，并将其划分为硬约束与软约束。针对硬约束，我们采用Qwen2.5-72B-Instruct生成对应的验证Python代码脚本。对于每条指令，我们从6种不同模型中随机采样一条回复，涵盖的模型包括：Llama3.1-8B-Instruct、Llama-3.3-70B-Instruct、Qwen2.5-7B-Instruct、Qwen2.5-72B-Instruct、QwQ-32B、DeepSeek-R1-Distilled-Qwen-32B。随后我们使用QwQ-32B为每条指令-回复对生成分步验证内容，用于判断该回复是否符合指令要求。最终我们共收集到约13万条指令-回复对及对应的分步验证内容。

如需了解更多细节，请参阅我们的论文及GitHub代码仓库（https://github.com/THU-KEG/VerIF）。

## 引用格式
bibtex
@misc{peng2025verif,
title={VerIF: 面向指令遵循任务强化学习的验证工程},
author={Hao Peng and Yunjia Qi and Xiaozhi Wang and Bin Xu and Lei Hou and Juanzi Li},
year={2025},
eprint={2506.09942},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.09942},
}

## 数据集卡片联系方式
如有任何疑问，请联系邮箱：[e-mail]