novastar112/toy_maze_2d_easy_allstep_thinking_future_rollout_cot_500k

该数据集名为ToyMaze2D Easy All-Step Future-Rollout COT，是基于VisGym ToyMaze2D的maze_2d/easy环境生成的。它用于图像到文本和强化学习任务，包含500,000行训练数据和100行测试数据，均为gzip压缩的JSONL格式。每行数据代表一个完整的轨迹对话，其中每个用户回合包括提示和JPEG图像项（包含image_prev、image和image_next字段），并验证了image_prev与image相等，最后一步的image_next与image相等。数据集中，每个轨迹有两个非停止移动步骤，提供四个一步未来展开图像（右、上、左、下）且loss=true；其他非停止步骤使用固定的简单思考模板；最终停止步骤使用固定的停止COT并输出(stop, stop)。提示步骤预算为20步，用户提示中显示剩余步数（max_steps - step_index - 1）。This dataset, named ToyMaze2D Easy All-Step Future-Rollout COT, is generated from the local VisGym ToyMaze2D maze_2d/easy environment. It is designed for image-to-text and reinforcement-learning tasks, containing 500,000 gzip-compressed JSONL rows for training and 100 for testing. Each row represents a full trajectory conversation, where every user turn includes a prompt and JPEG image items with image_prev, image, and image_next; it is validated that image_prev == image for every step, and the final step has image_next == image. Two non-stop move steps per trajectory contain four one-step future rollout images (right, up, left, down) with loss=true. Other non-stop steps use a fixed trivial thinking template, and the final stop step uses a fixed stop COT and outputs (stop, stop). The prompt step budget is 20 total steps for easy, with the user prompt indicating max_steps - step_index - 1 remaining steps.