Questions about replicating NLU experiments

[lky-violet]

Dear Author, Hello, I encountered the following issues while reproducing the NLU-deberta-v2-xxlarge experiment in the example folder:

1. After running deberta_v2_xxlarge_mnli.sh file, I find it does not generate lora file such as deberta_v2_xxlarge_lora_mnli.bin. However generate whole model parameter in mnli/model path. I wonder if there is something wrong with my operation？
2. In readme.md in NLU, the section "Evaluate the checkpoints" use do_eval parameter, however, should we use 'do_predict' instead?
3. After submitting the result file on kaggle-mnli challenge, the result of the lora finetune only get public score:0.33156. This is a big difference from the results reported in the paper. I wonder if this is because of the hyperparameter selection (I didn't change the provided hyperparameters) or because of the first problem mentioned above?