Training Instability with IQL: Loss Explodes After Initial Epochs

[shhmxy2]

Hi - Thanks for the great work! I'm experiencing training instability when running IQL on my custom dataset (~8000 samples). The loss decreases normally for the first 2-3 epochs but then suddenly increases and diverges. I've tried reducing learning rate from 3e-4 to 3e-5, but it only delays the issue. I wonder if anyone else has encountered similar instability? Any suggestions for hyperparameter tuning or potential implementation issues to check? Thanks!