We have a lot of config files that do different things which might be confusing to new users and lead to unexpected results. For example, the callback configs sometimes save a checkpoint every 25 epochs, sometimes they only save the last checkpoint. Sometimes, they save the best 3, sometimes only the best 1.

I tried to reduce the number of configs and find a consistent naming pattern. More precisely, I

• removed the weightings directory and loss/weighting_chebi100.yml - those were all static lists of weights which are not needed anymore since those weights get calculated on the fly by the weighted BCE.
• removed the metrics which had a specific label number attached. Those label numbers are automatically inferred from the dataset
• restored the default trainer and default callbacks to their original state (which means removing the roc-auc). This fixes Missing labels in loss kwargs causing issue for electra model / Issues from PR#130 #140 (comment). For other callbacks, the roc-auc is still used (e.g. binary_callbacks.yml and early_stop_callbacks_roc-auc.yml
• renamed the solCur and tox21 callbacks (as far as I can tell, they are not specific to the dataset)
• removed some comments (I assume that those parts are outdated)
• added an upper limit of 1 to the weighted BCE's beta parameter

solves #139