关于think的问题

[icey-zhang]

Prompt中回答要求包含think

数据集中没有关于关于think的回答
{"id": "amex-1220", "image_path": "/root/workspace/datasets/GUI-R1/pokeman_train_ORI/image/amex-1220.jpg", "instruction": "Open Citymapper. Find the nearest E-scooter station. How long does it take to walk there?", "gt_input_text": "UP", "gt_action": "scroll", "gt_bbox": [-0.1, -0.1], "history": "None", "task_type": "high"}

是只监督了action么？这样测试的时候也只是提高了action的准确性，不考虑think的回答么？

[icey-zhang]

另外想请问一下verl/utils/dataset.py代码中的prompt = prompt.replace("", "<|vision_start|><|image_pad|><|vision_end|>")为何要做这个替换？

[RainBowLuoCS]

奖励只有在最后有

[icey-zhang]

？没理解你说的是什么意思？“奖励只有在最后有“