Moonlight556/kimi-linear-48b-a3b-target-matched-math-240k

该数据集名为kimi-linear-48b-a3b-target-matched-math-240k，包含239,467行数学推理轨迹，这些轨迹是针对目标模型moonshotai/Kimi-Linear-48B-A3B-Instruct重新生成的。数据集专门用于训练DFlash推测解码的起草模型，并应用于la-draftery项目。target-matched表示用户提示来自Nemotron v2数学语料库，助手完成内容是目标模型自身的输出，即每个提示都发送给目标模型并捕获其完成内容，这使得起草模型能够准确预测目标模型将生成的令牌，从而在推测解码中实现高接受长度。数据模式包括ID、生成器、来源、对话内容（用户和助手）、思考字段（为空）和状态（全部为成功）。生成方法涉及使用SGLang服务器、OpenAI协议客户端、贪婪解码（温度=0）、最大令牌数为3072，且禁用思考功能。数据集可用于训练DFlash起草模型，并已证明在la-draftery Phase 1.2中实现5.7095 Math500平均接受长度和1.84倍SGLang spec-v2服务加速。许可证为Apache-2.0，与目标模型输出相同。The dataset kimi-linear-48b-a3b-target-matched-math-240k contains 239,467 rows of math-reasoning trajectories regenerated against the target model moonshotai/Kimi-Linear-48B-A3B-Instruct. It is used to train DFlash speculative-decoding drafters in la-draftery. Target-matched means the user prompts come from the Nemotron v2 math corpus, and the assistant completions are the target models own outputs—each prompt was sent to the target model and its completion was captured. Drafters trained on this data learn to predict exactly the tokens the target would emit, which is required for high acceptance length in speculative decoding. The schema includes ID, generator, source, conversations (user and assistant), thinking (null), and status (all success). Generation method uses SGLang server, OpenAI-protocol client with greedy decoding (T=0), max tokens 3072, and thinking disabled. The dataset is used for training DFlash drafters, and reproducible results show 5.7095 Math500 mean acceptance length and 1.84x SGLang spec-v2 speedup in la-draftery Phase 1.2. License is Apache-2.0, same as the target model outputs.