DSWorld-8K
收藏资源简介:
DSWorld-8K是由香港科技大学(广州)研究团队构建的高质量数据集,旨在支持数据科学世界模型的训练。该数据集包含8000条数据科学代理轨迹,涵盖真实世界和合成状态转换,每条样本包含当前状态、操作、下一状态及解释转换逻辑的思维链轨迹,数据来源基于MMTU中的60K真实表格和NumPy/Pandas生态系统操作库。数据集通过结合真实轨迹收集与可扩展合成流程构建,利用高级LLM生成多样工作流状态和数据科学操作,并执行验证以确保样本质量。该数据集主要应用于训练数据科学世界模型,以预测数据科学操作效果,解决自主数据科学代理因依赖试错计算而导致的效率瓶颈问题,提升代理训练和推理速度。
DSWorld-8K is a high-quality dataset constructed by the research team from The Hong Kong University of Science and Technology (Guangzhou), designed to support the training of data science world models. It contains 8,000 data science agent trajectories covering both real-world and synthetic state transitions. Each sample includes the current state, action, next state, and a chain-of-thought trajectory that explicates the transition logic. The data sources are based on 60K real tables from MMTU and the NumPy/Pandas ecosystem operation libraries. The dataset is built by combining real trajectory collection and scalable synthetic workflows, leveraging advanced LLMs to generate diverse workflow states and data science operations, and conducting validation to ensure sample quality. This dataset is primarily used for training data science world models to predict the effects of data science operations, addressing the efficiency bottlenecks of autonomous data science agents caused by relying on trial-and-error computations, and improving the training and inference speeds of these agents.




