AgentEvol-7B has too short context length; Training scripts are not the same as in the paper.

[nntoan209]

Hi, can you please explain these problems:

• The training scripts are not complete. In the paper, you stated that there are 2 training phases: Behavioral cloning (BC) and AgentEvol, which is a method to investigate the potential of agent self-evolution. However, the provided codes for the 2 training phases seem to be the same, I think the AgentEvol code is not the same as described in the paper.

• How is the model AgentGym/AgentEvol-7B evaluated? This model has a context length of 2048, and when I tried to evaluate this model on some environments, it failed because the context length is too short. Could you provide the evaluation code to reproduce the reported results?

[lzh1998-jansen]

同问，如何复现这个分数

[SHIFTTTTTTTT]

同问，甚至代码都跑不了，会卡住