Zigeng/dParallel_Dream_Distill_Data
收藏Hugging Face2025-10-14 更新2025-10-25 收录
下载链接：
https://hf-mirror.com/datasets/Zigeng/dParallel_Dream_Distill_Data
下载链接
链接失效反馈官方服务：
资源简介：
dParallel-Dream-Distill数据集是用于确定性强化蒸馏过程的训练数据集，其中包含了由LLaDA-8B-Instruct和Dream-7B-Instruct模型根据公开数据集提示生成的响应。LLaDA模型使用GSM8K、PRM12K和Numina-Math数据集的提示，而Dream模型则在此基础上增加了AceCode数据集的子集提示。所有数据均由模型自身生成，没有使用外部数据作为目标。
The dParallel-Dream-Distill dataset is a training dataset for the certainty-forcing distillation process, containing responses generated by the LLaDA-8B-Instruct and Dream-7B-Instruct models based on prompts from public datasets. The LLaDA model uses prompts from GSM8K, PRM12K, and the Numina-Math dataset, while the Dream model adds prompts from a subset of the AceCode dataset. All data are generated by the model itself without using external data as targets.
提供机构：
Zigeng
