Sampling strategy for FewShotClassifier

[BenjaminBossan]

We just released the FewShotClassifier class in #979.

One feature we thought could be interesting to implement is a way to change the sampling strategy for the few shot samples, i.e. how the samples are chosen for few-shot learning.

Right now, the sampling is hard-coded and basically tries to add each label at least once. This seems reasonable but there are situations where other strategies could make sense. Therefore, I would like to see a feature that allows setting the sampling strategy as a parameter. Options that come to mind:

• Stratified sampling: roughly what we have now, but not quite
• Fully random sampling: sample regardless of label
• Similarity-based sampling: use the current sample to find similar samples from the training data
• Custom sampling: Allow users to pass a callable that performs the sampling

[BenjaminBossan]

Closing this in favor of #989, which collects all TODOs in relation to LLM classification.