WithinUsAI/Meta_Muse_Spark_Distilled_5k

Meta Muse Spark Distilled — 5K Reasoning Traces是一个合成蒸馏数据集，设计于2026年5月，旨在模拟Meta前沿模型Muse Spark（2026）的思维和推理风格。它并非直接复制Muse Spark的输出，而是通过程序化生成5,000个独特示例，以教授逐步推理过程：理解→计划→执行→验证。数据集采用JSON Lines格式，每个示例包含唯一标识符、类别（如算术、代数、逻辑等）、用户指令、显式的逐步推理轨迹和最终响应。涵盖10个类别，每个类别500个示例，无重复。设计原则包括问题重述、显式规划、展示中间工作、自我验证，并保持平衡、事实性的语气。该数据集可用于基本微调（如Llama、Mistral等模型），通过Hugging Face加载，建议在训练中混合70%的本数据集（推理过程）和30%的领域任务（无轨迹），以帮助模型内化推理风格。限制包括数据为合成、简化数学/逻辑以覆盖广度而非深度，不包含专有知识或受版权文本，适用于研究和模型行为塑造，而非复制封闭模型权重。许可证允许用于训练开放模型，并要求在发布训练模型时引用。Meta Muse Spark Distilled — 5K Reasoning Traces is a synthetic distillation dataset designed to mirror the thinking and reasoning style of Metas newest frontier model, Muse Spark (2026). Created May 2026, it is NOT copied output from Muse Spark; all examples are programmatically generated to teach step-by-step reasoning: Understand → Plan → Execute → Verify. The dataset contains 5,000 unique examples in JSON Lines format, each with fields: id, category (e.g., arithmetic, algebra, logic), instruction, thinking_trace (explicit step-by-step reasoning), and response. It covers 10 categories with 500 examples each, no duplicates. Design principles include starting with problem restatement, explicit planning, showing intermediate work, self-verification, balanced factual tone, and conciseness. It can be used for basic fine-tuning (e.g., Llama 3/4, Mistral) via Hugging Face, with a suggested training mix of 70% this dataset (reasoning process) and 30% domain tasks (without traces) to teach the model to internalize the trace style. Limitations: synthetic data, simplified math/logic for breadth not depth, no proprietary knowledge or copyrighted text, for research and model behavior shaping, not for replicating closed model weights. License allows use, modification, and distribution for training open models, with citation required if publishing trained models.