Celestia2

**Celestia 2** is a multi-turn agent-instruct dataset containing science data.

This dataset focuses on challenging multi-turn conversations and contains:

- 176k rows of synthetic multi-turn science-instruct data, using [Microsoft's AgentInstruct style.](https://huggingface.co/datasets/microsoft/orca-agentinstruct-1M-v1) All prompts and responses are synthetically generated using [Llama 3.1 405b Instruct.](https://huggingface.co/meta-llama/Meta-Llama-3.1-405B-Instruct) Primary subjects are physics, chemistry, biology, and computer science; secondary subjects include Earth science, astronomy, and information theory.

This dataset contains synthetically generated data and has not been subject to manual review.

**Celestia 2** 是一款包含科学数据的多轮智能体指令（agent-instruct）数据集。本数据集聚焦于具有挑战性的多轮对话场景，具体包含以下内容：
- 共计17.6万条合成多轮科学指令数据，采用微软（Microsoft）的AgentInstruct风格，相关风格数据集可参考链接：https://huggingface.co/datasets/microsoft/orca-agentinstruct-1M-v1。所有提示词与回复均通过Llama 3.1 405B Instruct模型合成生成，对应模型的官方链接为：https://huggingface.co/meta-llama/Meta-Llama-3.1-405B-Instruct。数据集的核心学科涵盖物理学、化学、生物学与计算机科学，次要学科包括地球科学、天文学与信息论。
本数据集仅包含合成生成的数据，未经过人工审核。