qwen36-distill-opus-dsv4
收藏资源简介：
Qwen3.6蒸馏数据集（Opus + DeepSeek v4）是一个专门为知识蒸馏任务设计的文本生成数据集，旨在将Claude Opus模型的高级推理轨迹和DeepSeek v4模型的智能体思维过程蒸馏到Qwen3.6 27B模型中，以提升其推理和规划能力。数据集由五个开源数据集（来自TeichAI和Jackrong的系列数据集）经过预处理、归一化和整合而成，格式适配Qwen3.6的聊天模板，并在助手角色的最后一条消息中引入了特殊的reasoning_content字段，用于存储模型的内部推理或思考过程。数据集采用课程学习策略组织，分为三个阶段：第一阶段（格式归纳）包含11,621个训练样本和581个测试样本；第二阶段（深度推理）包含1,630个训练样本和81个测试样本；第三阶段（智能体思维轨迹）包含890个训练样本和44个测试样本。总计提供约13,000个训练样本和约700个测试样本。在构建过程中，原始输入的23,042条数据经过严格过滤，保留了18,801条（通过率81.6%），过滤标准包括剔除模拟思考、长度不佳、字段为空或缺失、内容与思考完全重复等不合格样本。数据集适用于大语言模型的推理能力增强、思维链蒸馏、智能体行为模仿等任务。整体遵循Apache-2.0许可证，但各源数据集保留其原始许可证。
The Qwen3.6 distillation dataset (Opus + DeepSeek v4) is a text generation dataset specifically designed for knowledge distillation tasks. Its core purpose is to distill the advanced reasoning trajectories from the Claude Opus model and the agent thought processes from the DeepSeek v4 model into the Qwen3.6 27B model, to enhance the latters reasoning and planning capabilities. The dataset is constructed by preprocessing, normalizing, and integrating five open-source datasets (series datasets provided by TeichAI and Jackrong), with the final format adapted to Qwen3.6s chat template, and introduces a special `reasoning_content` field in the last message of the assistant role to store the models internal reasoning or thought processes. The dataset is organized using a curriculum learning strategy, divided into three stages: the first stage (format induction) contains 11,621 training samples and 581 test samples; the second stage (deep reasoning) contains 1,630 training samples and 81 test samples; the third stage (agent thought trajectories) contains 890 training samples and 44 test samples. In total, it provides approximately 13,000 training samples and about 700 test samples. During construction, 23,042 original input data points were strictly filtered, resulting in 18,801 retained (81.6% pass rate). Filtering criteria include removing unqualified samples such as simulated thoughts, poor length, empty or missing fields, and content completely duplicating thoughts. The dataset is suitable for tasks like large language model reasoning enhancement, chain-of-thought distillation, and agent behavior imitation. The overall dataset follows the Apache-2.0 license, but each source dataset retains its original license.
