Finetune a model for Multi-label classification/regression - matrix/vactor as a label?


I have my own dataset that is comprised of hundreds of sentences and each sentence has 10 different sentiment labels and a discrete score for each of the labels (from 1-10). So essentially, my ith input for the training the model looks something like sen_i ; a matrix size (1,10*10) (or 10, 10), assuming I avoid a regression by one-hot encoding the discrete scores for each label (which is another question - is that correct to take such an approach?) I was wondering if there’s any model on the hub I could finetune that could learn from a label that is a matrix/vector and not just a scalar.

Thank you.