CPO & SimPO Loss

[psinger]

This PR implements new CPOLoss and SimPOLoss

Please review after #753

Closes #732

[pascal-pfeiffer]

Thank you for the integration @psinger !

It seems to work nicely, though default values give bad results.
We might want to consider not to rename the Loss functions to avoid issues when starting from older experiments.
Also, there is a minor issue with the calculation of the number of experiments being started in grid search.

[psinger]

It seems to work nicely, though default values give bad results.

Bad in what way? It is based on best experiments I got for them, but it might be more sensitive depending on what model / data one uses.

[pascal-pfeiffer]

It seems to work nicely, though default values give bad results.

Bad in what way? It is based on best experiments I got for them, but it might be more sensitive depending on what model / data one uses.

Just running the default experiment with Loss function changed to SPO gives for example these predictions:

But it is probably fine, it is expected that these things need tuning.

[pascal-pfeiffer]

Thank you @psinger , lgtm!