Custom loss function

[mariiapronesti01]

Hi! I am trying to finetune Llama3.1 3B and I would like to use a customized loss function.
I read in a past issue that I have to remove the causal LM head and replace it with mine. Since I am not an expert, can I ask for more details and information?

Thanks a lot and congrats for the great library!

[danielhanchen]

It's best to ask on Discord for this :) Generally it depends on why you're trying to make a custom loss - is it for classification? If yes, then we're actively working on making AutoModelForSequenceClassification work - hopefully by next week

[mariiapronesti01]

I'll join the channel then!
No it's not classification. I am working with graphs and I would like to use a custom metric that measures the distance between two graphs as loss function

[Gladiator07]

It's best to ask on Discord for this :) Generally it depends on why you're trying to make a custom loss - is it for classification? If yes, then we're actively working on making AutoModelForSequenceClassification work - hopefully by next week

Awesome, much awaited, will there be a announcement post ? This will be a really nice addition!

Also will it be possible to init the model with num_labels as in the huggingface's implementation then subclass Trainer and override the compute_loss function so that the same architecture can be used in reward model setting (bradley terry loss), etc ?