Fix embedding-based classification documentation

[SebastianZettAA]

The doc string is correct for class EmbeddingBasedClassify (Scores will be between 0 and 1 but do not have to add up to one.), but not in the classification example notebook classification.ipynb.

Two followup questions here:

• The prompt-based classification is denoted as single-label classification, whereas the embedding-based classification as multi-label. I do not see a real reason why this differentiation is made, in my opinion both approaches (log-prob scores vs. cosine similarity scores) could be used for both classification tasks (i.e., only assign the class with the highest score, vs. assigning multiple classes surpassing a to be defined threshold).
• Is there a reason why the prompt-based classification scores are normalized (to sum up to 1), but those of embedding-based classification not?

[MerlinKallenbornTNG]

Aloha,
what's the status of this PR? Looks like nothing happened since two weeks. Can it be merged or has it become obsolete?

[SebastianZettAA]

I opened the PR when trying to use the classification task(s) for a client project when I still was on the customer team. At that time I did not feel responsible nor entitled to work more directly on the IL than asking above questions.
Since I was told @NickyHavoc primarily worked on this topic I suggest we talk today or tomorrow about this in person?