hkust-nlp/drkernel-coldstart-8k

DR.Kernel冷启动数据集用于在强化学习（RL）之前进行监督微调（SFT），以初始化Triton代码生成和迭代优化的能力。数据集包含8,920条多轮对话轨迹，每条轨迹包含10条消息，固定角色顺序。数据以Parquet格式存储，包含多个字段，如messages、uuid、entry_point等。数据集的使用方法包括加载数据集和进行SFT训练。数据集的收集过程涉及从强大的专有教师模型中提取多轮交互数据，并在每轮对话中添加执行反馈以促进迭代优化。数据集的相关资源包括论文、代码库、训练数据、验证数据等。

The DR.Kernel Cold-Start Dataset is used for supervised fine-tuning (SFT) before reinforcement learning (RL) to initialize the ability of Triton code generation and iterative optimization. The dataset contains 8,920 multi-turn trajectories, each consisting of 10 messages with a fixed role order. The data is stored in Parquet format and includes multiple fields such as messages, uuid, entry_point, etc. The usage of the dataset includes loading the dataset and performing SFT training. The data collection process involves distilling multi-turn interaction data from strong proprietary teachers and appending execution feedback in each turn to prompt iterative refinement. Related resources of the dataset include papers, code repositories, training data, validation data, etc.