TeamClaude/sft-100k-mixture

SFT 100K Mixture数据集是一个经过精心策划、包含100,000行数据的监督微调（SFT）混合数据集，专门用于大型语言模型（LLM）的指令调优。该数据集由三个主要组件组成：逻辑核心（占60%，来源为PrimeIntellect/SYNTHETIC-2-SFT-verified，包含60,000行数据，专注于推理、数学和编码任务，并带有经过验证的DeepSeek-R1轨迹）、对话粘合剂（占30%，来源为Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered，包含30,000行数据，旨在提升多轮对话的流畅性和语气对齐）以及人类元素（占10%，来源为allenai/WildChat-4.8M，包含10,000行数据，提供真实世界的英语、无毒提示）。数据集采用统一的OpenAI messages格式，每个样本包括一个消息列表（包含角色和内容字段）和一个来源标识。它适用于文本生成任务，主要用于英语环境，并带有SFT、指令调优和混合等标签。用户可以通过Hugging Face的datasets库轻松加载和使用该数据集。

SFT 100K Mixture Dataset: A curated 100,000-row SFT mixture dataset for instruction-tuning LLMs. It consists of three components: Logic Core (60% from PrimeIntellect/SYNTHETIC-2-SFT-verified with 60,000 rows for reasoning, math, coding with verified DeepSeek-R1 traces), Conversational Glue (30% from Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered with 30,000 rows for multi-turn fluidity and tone alignment), and Human Element (10% from allenai/WildChat-4.8M with 10,000 rows for real-world English, non-toxic prompts). The dataset uses a unified OpenAI messages format with messages (role and content) and source fields. It is designed for text-generation tasks in English and tagged with SFT, instruction-tuning, and mixture. Usage involves loading via Hugging Face datasets.