zjunlp/KnowRL-Train-Data

KnowRL训练数据集是用于研究论文《KnowRL: 探索知识增强的强化学习以实现事实性》的训练数据集。该数据集旨在帮助大型语言模型（尤其是慢思考模型）识别其知识边界，减少幻觉现象，通过将外部知识整合到强化学习过程中，引导模型进行基于事实的慢思考。数据集包含三个核心JSON文件，分别对应KnowRL训练框架的不同阶段：冷启动监督微调数据、知识增强的强化学习训练数据和带有知识依据的强化学习训练数据。

The KnowRL training dataset is for the research paper KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality. This dataset aims to help large language models, particularly slow-thinking models, recognize their knowledge boundaries to reduce hallucinations, by integrating external knowledge into the reinforcement learning process, guiding the model to perform fact-based slow thinking. The dataset consists of three core JSON files, each corresponding to a different stage of the KnowRL training framework: cold-start supervised fine-tuning data, knowledgeable reinforcement learning training data, and reinforcement learning training data with grounding knowledge.