Llama Adapter finetuning parameters

[LCTPalmer]

Hi

Thank you so much for opening your data and code, it is very useful. I am trying to implement your method as a baseline, but want to be sure of training hyperparameters that you use. The exps/finetune.sh code includes some hard-coded hyperparameters (the same as the orinignal llama-adapter repo) - are you using these for the results in your paper? And if so are you training on the full training set in one go or using some other strategies? Any help appreciated.

Best regards
Luke

[ChonghaoSima]

In the paper we did not provide experiments of llama-adapter on DriveLM. This (the llama-adapter code) is provided for the DriveLM challenge as a starter-code. We are using the full training set to train DriveLM-Agent, and more detail (training strategies) are in the paper.

[LCTPalmer]

Thanks for your respnse. In the paper appendix section F.5 and Table 13 you mention training Llama-Adapter V2 and report results; I was hoping to benchmark against planning metric locally against your method so would be great to have those hyperparams (as I am adding some logic to llama-adapter in hopes of general improvement). Alternatively, is there an entry in the leaderboard for vanilla Llama-AdapterV2 I could reference?

Best regards
Luke