VerInstruct

# 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
- **Source:** This data is sourced from [Crab](https://huggingface.co/datasets/THU-KEG/Crab-SFT), and we add verification signals for each instance.

## Uses

This data is used for RL training 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>,
"prompt": <prompt>,
"checkers": <constraints in the prompt to check>,
"functions": <verification function for each checker>
}

```
## Dataset Creation

### Source Data

This data is sourced from [Crab](https://huggingface.co/datasets/THU-KEG/Crab-SFT), and we add verification signals for each instance.

#### Data Collection and Processing

For each instruction, we annotate the verification methods for the constraints in the instruction.
For hard constraints, including format, keyward, and length, we use Qwen2.5-72B-Instruct to generate the verification code.
For soft constraints, we adpot an advanced reasoning LLM as the LLM verifier. We recommand use QwQ, and we also distill a small-but-effective 7B [verifier](https://huggingface.co/THU-KEG/IF-Verifier-7B).

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：数据集名称

## 数据集详情

### 数据集描述

<!-- 请提供该数据集的详细概述 -->

- **整理者：** 郝鹏@THUKEG
- **涉及语言（自然语言处理）：** 英语、中文
- **授权协议：** 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
- **数据来源：** 本数据集源自[Crab](https://huggingface.co/datasets/THU-KEG/Crab-SFT)，并为每条数据实例添加了验证信号。

## 数据集用途

本数据集用于指令跟随任务的强化学习（Reinforcement Learning, RL）训练。

## 数据集结构

数据集采用`jsonl`格式存储，每一行均为一条JSON数据项，格式如下：

{
"id": <数据编号>,
"prompt": <指令提示>,
"checkers": <需校验的指令约束>,
"functions": <针对各校验项的验证函数>
}

## 数据集构建

### 源数据

本数据集源自[Crab](https://huggingface.co/datasets/THU-KEG/Crab-SFT)，并为每条数据实例添加了验证信号。

#### 数据收集与处理流程

针对每条指令，我们为指令中的约束项标注了验证方法。
对于格式、关键词、长度等硬约束，我们使用Qwen2.5-72B-Instruct生成验证代码。
对于软约束，我们采用高级推理大语言模型（Large Language Model, LLM）作为验证器。我们推荐使用QwQ，同时还蒸馏得到了一款小巧高效的7B参数[验证模型](https://huggingface.co/THU-KEG/IF-Verifier-7B)。

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

@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},
}

## 数据集卡片联络方式

如有任何疑问，请联系[[e-mail]]。