FreedomIntelligence/AceGPT-v2-AlignmentData
收藏Hugging Face2025-12-29 更新2024-12-14 收录
下载链接：
https://hf-mirror.com/datasets/FreedomIntelligence/AceGPT-v2-AlignmentData
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是为了在AceGPT-v2中实现原生对齐而构建的。通过从ArabicText 2022中随机选择96K样本，并使用GPT-4-turbo根据提供的提示重写这些数据，最终组织成对以创建对齐LLM的训练数据。重写过程遵循特定的标准，包括语法和句法、文化适宜性、噪音去除、一致性、数学公式格式化和代码格式化等。
This dataset is designed to achieve native alignment in AceGPT-v2 by training a small alignment model to filter the entire pre-train dataset. It consists of 96K randomly selected samples from ArabicText 2022, rewritten using GPT-4-turbo according to specific prompts, and organized into pairs to create training data for the Alignment LLM. The rewriting process adheres to criteria such as grammar and syntax correction, cultural appropriateness, noise removal, consistency, and standardization of mathematical formulas and code formatting.
提供机构：
FreedomIntelligence
