You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
I find that the quantisation losses are higher for GPTJ than LLama which seems to stay pretty low.
This is with a sample size of 2048 * 2 on C4 dataset. I find it gets better the larger I make the dataset.
I also notice that as you go through the layers it gets worse. So the loss is highest for layer 28/28 (approximately).
Why is the avg loss so high for GPTJ compare to LLama? With Llama I am also able to use a much smaller dataset: like 1024 to achieve lower loss.
Beta Was this translation helpful? Give feedback.
All reactions