Havoc999/The-Claude-Dataset

该数据集是一个大规模、经过清理和统一处理的集合，包含624,252个推理轨迹，这些轨迹蒸馏自Claude 3.5 Sonnet和Opus（版本4.6/4.7）。它专门设计用于训练小型高性能模型（例如nanowhale-100m），这些模型采用多头部潜在注意力（MLA）或混合专家（MoE）架构，其中高密度思维链数据至关重要。数据集通过聚合五个高质量蒸馏源（包括纯推理轨迹、大规模指令遵循和过滤推理样本）构建，并解决了模式规范化、箭头对齐和清理等技术壁垒，确保数据格式统一为OpenAI/HuggingFace消息格式。数据以Parquet格式提供，总行数为624,252，优化用于MLA/MoE架构的训练。This dataset is a massive, cleaned, and unified collection of 624,252 reasoning traces distilled from Claude 3.5 Sonnet and Opus (versions 4.6/4.7). It was specifically engineered to train small, high-performance models (like nanowhale-100m) that utilize Multi-Head Latent Attention (MLA) or Mixture of Experts (MoE) architectures, where high-density Chain-of-Thought (CoT) data is critical. The dataset aggregates five high-quality distillation sources, including pure reasoning traces, large-scale instruction following, and filtered reasoning samples, and addresses technical barriers such as schema normalization, arrow alignment, and cleaning to ensure a unified OpenAI/HuggingFace Message format. It is provided in Parquet format with 624,252 rows and optimized for training MLA/MoE architectures.