Randomness of the output of the trained model

#68
by Sam1989 - opened

I optimized a Gemma-2b model for a classification task in 4-bit quantization with LoRA. It shows remarkably good results on test data. The model has been saved. When I reload the model and make predictions, I notice that the probabilities for the classes change slightly after each prediction. I let the model predict the same text 5 times in a row and got the following values ​​for the best class:

[0.778,0.746,0.732,0.693,0.712]

Where does this randomness come from? It should be the same every prediction ? Because the weights and biases of the model are no longer changed .

Google org
edited Oct 17

Hi @Sam1989 , Sorry for the late response, Could you please provide some more details on the issue like the reproducible code to replicate the issue for better understanding. Thank you.

Sign up or log in to comment