TESS-Computer/minecraft-vla-stage1

该数据集包含来自Minecraft游戏的帧-动作对，专为遵循Lumine方法论的VLA模型训练而设计。数据集基于OpenAI VPT承包商数据的7.x子集，包含约17,886个视频(约330小时的早期游戏内容)，聚焦于新世界创建后的前30分钟游戏内容。每个样本包含640x360 JPEG图像帧、视频ID、帧索引和Lumine格式的动作字符串。动作字符串包含鼠标移动、滚动(始终为0)和按键组合信息。数据集经过处理，帧率从20FPS降采样至5FPS，移除了空闲帧和加载画面。这是三阶段训练流程的第一阶段，专注于动作预训练(学习观察→动作映射)。

This dataset contains frame-action pairs from Minecraft gameplay, designed for training VLA models following the Lumine methodology. Processed from OpenAIs VPT contractor dataset (7.x subset), it includes ~17,886 videos (~330 hours of early-game gameplay) focusing on the first 30 minutes of new worlds. Each sample contains a 640x360 JPEG frame, video identifier, frame index, and Lumine-format action string. The action string includes mouse movements, scroll (always 0), and key combinations. The data is processed to 5 FPS (downsampled from VPTs 20 FPS) with idle frames and loading screens filtered out. This is Stage 1 of a 3-stage training pipeline focused on action pretraining (learning observation→action mapping).