drkernel-coldstart-8k
收藏资源简介:
DR.Kernel冷启动数据集用于在DR.Kernel中进行监督微调(SFT)以初始化内核生成能力(Triton编码+迭代优化),然后再进行TRLOO/MRS/PR/PRS强化学习。数据集包含8,920条多轮对话轨迹,每条轨迹由5轮对话组成,共10条消息,角色顺序固定为用户和助理交替。数据以Parquet表格形式存储,包含字段如messages(完整的多轮聊天历史)、uuid(样本ID)、entry_point(入口类/函数名)、final_speedup(最终加速比)等。数据集通过KernelGYM反馈收集,每轮对话包含服务器反馈和改进请求,支持多轮SFT训练。数据来源于ByteDance-Seed/cudaLLM-data,并在此基础上构建了多轮轨迹(迭代反馈+改进)。使用该数据集时,请引用DR.Kernel和原始查询源。
The DR.Kernel Cold-start Dataset is intended for supervised fine-tuning (SFT) within the DR.Kernel framework to initialize the kernel generation capability (Triton coding + iterative optimization), prior to reinforcement learning with TRLOO/MRS/PR/PRS. The dataset comprises 8,920 multi-turn dialogue trajectories, each consisting of 5 conversation rounds with a total of 10 messages, where the speaker sequence strictly alternates between user and assistant. The data is stored in Parquet table format, with fields including messages (full multi-round chat history), uuid (sample ID), entry_point (entry class/function name), final_speedup (final speedup ratio), and more. This dataset is collected via KernelGYM feedback, where each conversation round includes server feedback and improvement requests, enabling multi-turn SFT training. The dataset is derived from ByteDance-Seed/cudaLLM-data, and multi-turn trajectories (iterative feedback + improvement) are constructed based on this source. When using this dataset, please cite DR.Kernel and the original query source.




