NiuTrans/GRAM-pre-training-566k

这是一个用于预训练GRAM模型的dataset。数据集的格式包括指导语、输入以及两个助手的回答输出。该数据集是从llm-blender/Unified-Feedback数据集中筛选而来，去除了过长的数据和包含乱码的数据。数据集与强化学习中的奖励泛化和偏好相关。

This is a dataset for pre-training the GRAM model. The dataset format includes an instruction, input, and outputs from two assistants. The dataset is sourced from the llm-blender/Unified-Feedback dataset, filtered for length and character issues. The dataset is related to reward generalization and preference in reinforcement learning.