Overtrained?

#2
by exquisiteart - opened

I think LORA might be a bit overtrained, I have had had 70B 3.1 model get stuck in a loop repeating pretty much the same thing over and over again with minute variations. What dataset did you use?

yes, it was overtrained, i noticed that, im gonna update the lora, but as you can see, its uncensored, i trained in 3 epochs, can you tell me how much it should be to not be overtrained? if you say to me, im gonna train again after you answer me.

arafatar/toxic_uncensored_LGBTQ_csv this is the dataset that i used

dont worry, i found the problem, im gonna train again now

im testing a new version now, the logic seems to be more upgraded but it repeats itself after 4 messages, i think its the way i trained, im gonna modify my dataset, dont worry.

Guilherme34 changed discussion status to closed

Makes sense, I'm generating some training data on my own as well. How long does it take you to run through this? I do have some mid level multi-gpu hardware available

its definetely the Way i managed the dataset, if you add more examples in one input to train its gonna work better.

Sign up or log in to comment