smh13171317/NOESIS-1M-reasoning-router-code-math-psych-opus47-deepseek4-qwen36-gemini31-r1-gpt54
收藏资源简介：
NOESIS多语言推理路由器监督微调数据集（包含100万条全量数据和5万条精选数据）是一个大规模多语言指令微调数据集，专门为NOESIS QwQ+DeepSeek-R1混合专家（MoE）管道构建。数据集包含约105万条记录，覆盖30多种语言，重点支持英语、中文、印地语、西班牙语、阿拉伯语、法语、俄语等主要语言。数据内容涵盖通用指令、代码生成、数学推理、心理学问题以及带有思维链（chain-of-thought）的推理轨迹。数据集采用JSONL格式，每条记录包含用户问题和助手回答的对话文本。数据来源多样，包括Aya多语言指令数据集、Claude Sonnet 4.6的高质量英文助手回复、DeepSeek-R1的推理轨迹蒸馏、NOESIS翻译对以及其他混合来源。精选的5万条子集采用质量评分策略优先选择包含推理轨迹、长回复、代码和数学内容的高质量样本，并按语言配额平衡分布。该数据集主要用于：1）DoRA（Difference-based Low-Rank Adaptation）监督微调；2）混合专家模型的路由器权重训练；3）多语言指令调优与推理轨迹蒸馏。目标模型架构为NOESIS-QwQ-R1管道（QwQ-32B与DeepSeek-R1-32B的TIES合并后形成的16专家条件混合专家模型）。数据集遵循Apache 2.0许可证，部分数据衍生自CoForAI的Aya数据集。
The NOESIS Multilingual Reasoning Router SFT Dataset (1M + 50K curated) is a large-scale multilingual supervised fine-tuning dataset built for the NOESIS QwQ+DeepSeek-R1 Mixture of Experts (MoE) pipeline. It contains approximately 1.05 million records covering over 30 languages, with emphasis on English, Chinese, Hindi, Spanish, Arabic, French, Russian, and other major languages. The dataset encompasses general instruction, code generation, mathematical reasoning, psychology-related questions, and reasoning traces with chain-of-thought. Data is formatted in JSONL, with each record containing dialogue text in User: <question> Assistant: <answer> format. Sources include the Aya multilingual instruction dataset, high-quality English assistant turns from Claude Sonnet 4.6, reasoning traces distilled from DeepSeek-R1, NOESIS translation pairs, and other mixed SFT sources. The curated 50K subset employs a quality scoring strategy prioritizing samples with reasoning traces, long responses, code, and mathematical content, balanced across language quotas. The dataset is intended for: 1) DoRA (Difference-based Low-Rank Adaptation) supervised fine-tuning; 2) router (gate.weight) training for Conditional MoE architectures; 3) multilingual instruction tuning with reasoning trace distillation. The primary target is the NOESIS-QwQ-R1 pipeline (a TIES merge of QwQ-32B and DeepSeek-R1-32B formed into a 16-expert Conditional MoE). Licensed under Apache 2.0, with portions derived from Coheres Aya dataset.
