opus-research/opus-thinking

Opus Thinking Dataset是一个小型思维链推理数据集，包含534个带有明确思考步骤的对话示例，用于微调Opus 1.5模型以进行显式推理。数据集教导模型在回答前进行推理，格式为JSON，包含用户和助手的对话消息以及类别信息。助手响应遵循特定的思考格式，包括内部推理和实际响应。数据集分为11个类别，如事实问题、数学问题、建议请求等。数据集由Google Gemini 2.5 Flash Thinking生成，经过后处理验证。使用该数据集微调Opus 1.5后，模型的响应连贯性和任务理解能力有所提升。但数据集存在规模小、合成生成、质量参差不齐等局限性。

The Opus Thinking Dataset is a small chain-of-thought reasoning dataset containing 534 examples of conversations with explicit thinking steps, used to fine-tune Opus 1.5 for explicit reasoning. The dataset teaches models to reason before responding, formatted in JSON with user and assistant message conversations and category information. Assistant responses follow a specific thinking format, including internal reasoning and actual responses. The dataset is divided into 11 categories, such as factual questions, math problems, advice requests, etc. The dataset was generated by Google Gemini 2.5 Flash Thinking and validated through post-processing. Fine-tuning Opus 1.5 with this dataset improved the models response coherence and task understanding. However, the dataset has limitations such as small size, synthetic generation, and varying quality.