osunlp/early-experience
收藏资源简介:
该数据集是用于复现论文《Agent Learning via Early Experience》的监督微调数据,涵盖了8个不同的智能体环境(包括ALFWorld、WebShop、BFCL、TravelPlanner、TextCraft、AppWorld、ScienceWorld和WebArena)。每个环境提供三种训练范式的数据:模仿学习(IL,对应expert split)、自我反思(SR,对应expert和reflection splits组合)和隐式世界建模(IWM,对应iwm和expert splits组合)。数据采用ShareGPT格式(包含system、user和assistant消息),旨在通过早期经验提升智能体的学习性能。
This dataset serves as the supervised fine-tuning data for reproducing the paper *Agent Learning via Early Experience*. It covers 8 distinct agent environments, including ALFWorld, WebShop, BFCL, TravelPlanner, TextCraft, AppWorld, ScienceWorld, and WebArena. For each environment, data for three training paradigms are provided: Imitation Learning (IL, corresponding to the expert split), Self-Reflection (SR, corresponding to the combination of expert and reflection splits), and Implicit World Modeling (IWM, corresponding to the combination of IWM and expert splits). The data follows the ShareGPT format, which includes system, user, and assistant messages, and is designed to improve the learning performance of agents via early experience.




