GGOSinon/babyai-world-model-sft

BabyAI世界模型SFT数据集是用于BabyAI网格世界环境中世界模型的监督微调(SFT)的训练数据。世界模型学习预测给定当前状态和代理动作下的下一个观察和可用动作。每个示例都是一个环境转换，以聊天格式（系统/用户/助手）呈现：系统提示包含详细的环境规则和任务完成逻辑；用户输入包含目标、当前观察、可用动作和代理动作；助手输出则是预测的下一个观察和更新的可用动作（XML格式）。数据集包含58,238个训练示例和6,117个测试示例，来源轨迹来自RL训练滚动和基线评估。

Training data for supervised fine-tuning (SFT) of a world model for the BabyAI grid-world environment. The world model learns to predict the next observation and available actions given the current state and the agents action. Each example is a single environment transition in chat format (system/user/assistant): System prompt contains detailed environment rules and task completion logic; User input contains goal, current observation, available actions and agents action; Assistant output is the predicted next observation and updated available actions in XML format. The dataset contains 58,238 training examples and 6,117 test examples, sourced from RL training rollouts and baseline evaluations.