Does Nemo support soft targets or doing regression?

[m-harmonic]

Is your feature request related to a problem? Please describe.

I'm trying to do supervised learning given labeled data of an input token sequence mapping to a probability output. Does Nemo support a simple way to do this? For example, via soft targets (predicting one of two tokens each with a certain probability that sums to 1) as opposed to 1-hot vectors, or otherwise doing regression for a continuous value?

Describe the solution you'd like

Some modification to standard SFT training for example.

Describe alternatives you've considered

Have tried looking at distillation as a proxy, where the probabilities imitate the logits produced by a teacher model. But this is quite cumbersome.

Additional context