TESS-Computer/minecraft-vla-stage2

该数据集是Minecraft VLA（视觉-语言-动作）训练管道的第二阶段数据，专注于指令跟随训练。它在Stage 1的视觉运动数据基础上增加了任务指令，用于训练模型理解和执行指令。数据格式包括样本ID、视频ID、帧索引、任务指令、动作、任务类别、任务组、目标对象等字段。特别指出，指令仅在任务段的起始帧提供，延续帧的指令为空，要求模型保持目标上下文，这与推理时的使用方式一致。

This dataset adds task instructions to the Stage 1 visuomotor data, enabling instruction-following training. It is part of the Minecraft VLA (Vision-Language-Action) training pipeline. The data format includes fields such as sample ID, video ID, frame index, task instruction, action, task category, task group, target object, etc. Notably, instructions are provided only at the start of each task segment (is_segment_start=True), with continuation frames having empty instructions - requiring the model to maintain goal context, which matches inference usage.