LifelongAlignment/aifgen-piecewise-preference-shift

这是一个专门为终身对齐代理与强化学习而生的数据集子集，通过AIF-Gen框架生成，用于训练和基准测试大型语言模型上的RL方法。数据集包含三种不同风格的文本生成任务，针对政治话题，分别是模仿说唱歌手、模仿莎士比亚和正式风格。每个任务包含10,000个样本，总共90,000个样本，由gpt-4o-mini模型合成生成。该数据集遵循MIT许可，旨在用于大型语言模型上的静态连续/终身强化学习的基准测试。

This is a subset of datasets generated and curated for Lifelong Alignment of Agents with Reinforcement Learning, created using the AIF-Gen framework for training and benchmarking RL methods on Large Language models. The dataset includes text generation tasks for political topics in three different styles: like a rapper, like Shakespeare, and formal. Each task consists of 10,000 samples, totaling 90,000 samples, synthetically generated by the gpt-4o-mini model. The dataset is licensed under MIT and is intended for use in benchmarking static continual/lifelong Reinforcement Learning on Large Language models.