I had tried to use KE-T5 as a backbone but I think the model weights in PyTorch have some problems. So I recommend utilizing Kolang-T5.
-
python 3.7 or 3.8 can be used
pip install -r requirements.txt
I used CUDA 11.3 version
-
Wandb (Weights & Biases) is good logger which can view text logs
Update soon
bash scripts/pretrain_dialog.sh
python interact.py