THU-KEG/VerInstruct

这是一个用于指令遵循的强化学习训练的数据集，包含英文和中文两种语言的提示和验证信号。数据集从Crab数据集获取，并为每个实例添加了格式、关键字和长度等硬约束的验证方法以及软约束的验证。提供了大模型和小模型的验证器。

This is a dataset for reinforcement learning training in instruction following, containing prompts and verification signals in both English and Chinese. The dataset is sourced from the Crab dataset and verification methods for format, keyword, and length constraints, as well as soft constraints, are added for each instance. A large model and a small model verifier are provided.