Help in picking a model/head for labelling task

I want to fine-tune a model to give me a non-normalized set of scores for different labels based on an input text. Here are some more details:

For training:

  1. The input is a large string with 4 labels: happiness, sadness, anger, surprise
  2. each of the labels has a float value between 0-1 that is not normalized

I have a large dataset of these values that I would ideally use to fine-tune so that it can return a list of those values non-normalized.

I’m struggling to find an example that doesn’t only give probability scores. I would ideally like the model to output non-normalized scores for each of the labels. I’m also considering normalizing the input data, fine-tuning the model, and then de-normalizing the returned scores afterward, but I’m not sure if that’s plausible.

Thanks so much!