Fine tuning reducing randomness

Hi all :slight_smile:
Tried a couple configurations with fine tuning on Curie. In general, really cool results, but for some prompts, it just gets stuck on repeating the prompt. I increase temperature to the max, but to no avail.
Also, tried to avoid overfitting the data (don’t have a lot of it) so i just tuned 50 examples, one epoch, with 0,025 learning rate. Still stuck with repeating. Anyone got a way out?

Thanks a lot!

1 Like

Not the prompt, but sometimes I noticed the response gets stuck on repeat. I suspect it’s because I used a synthetic dataset for finetuning and did not include enough variety when making it.

It could also be your keyword. How are you ending your prompts? I do this:

[[original message]]


I do the training very consistently with two newlines followed by a very distinctive word that denotes the end of the input. Make sure you use the same exact format at inference time.

1 Like

Thanks! I actually use \n####\n in the end of prompt, and \n\n in the end of completion for the training… might be worth adding some stop sequence in the end of the completion maybe… following what you’re saying. I’m getting repetitions in the response as well (looping towards the end), but this is easier to take care of.

It’s best not to use whitespace at the end, and also the ##### is unclear, so you might want to switch to something more specific, or concrete. This is especially important if you end up making a finetune model with multiple purposes.

1 Like