MickeyFei
Results
3
comments of
MickeyFei
> For high-level tasks, the model predicts both "Thought: ..." and "Action: ...". For low-level tasks, the "Thought: ..." part is provided, and the model only needs to predict "Action:...
> All input data coordinates are normalized within the range of 0-1000 as relative coordinates, without involving absolute coordinate inputs. 抱歉，说错了，是相对坐标，请问是如何处理得到框对应的的相对坐标呢？
遇到了单论对话lora sft，loss 从step2 开始就为0的情况 8卡训练出现，单卡训练不出现