Sunshine-0921/eto-sft-trajectory

该数据集名为ETO-SFT-Trajectory，包含用于ETO（基于探索的轨迹优化）框架的专家轨迹。ETO是一个受人类学习试错过程启发的LLM代理学习框架，允许代理从失败轨迹中学习，并通过对比失败-成功轨迹对来优化策略。数据集覆盖三个代理环境：WebShop（基于人类演示和GPT-4探索，选择奖励大于0.7的轨迹）、ScienceWorld（使用启发式算法生成黄金轨迹）和ALFWorld（提供人类注释的轨迹用于模仿学习）。原始轨迹缺乏每个动作步骤的思维链（CoT）信息，因此利用GPT-4生成了相应的推理。数据集旨在支持代理的迭代策略学习，提升在未见场景下的泛化能力和任务解决效率。数据格式为JSON，包含ID和对话列表（人类和GPT角色之间的交互）。

This dataset, named ETO-SFT-Trajectory, contains expert trajectories for the ETO (Exploration-based Trajectory Optimization) framework. ETO is an LLM agent learning framework inspired by the trial and error process of human learning, enabling agents to learn from failure trajectories and optimize policies through contrastive failure-success trajectory pairs. The dataset covers three agent environments: WebShop (based on human demonstrations and GPT-4 exploration, selecting trajectories with rewards greater than 0.7), ScienceWorld (using heuristic algorithms to generate golden trajectories), and ALFWorld (providing human-annotated trajectories for imitation learning). The original trajectories lack Chain-of-Thought (CoT) information for each action step, so corresponding rationales are generated using GPT-4. The dataset aims to support iterative policy learning for agents, improving generalization in unseen scenarios and task-solving efficiency. The data format is JSON, including IDs and conversation lists (interactions between human and GPT roles).