1. Make number of rows to print a configurable parameter (default to 5).
2. Add parameter to enable index column printing. This implementation
   prints index of the original data frame if the passed data frame is
   derived.

Otherwise we don't know whether it improved the accuracy or not.

We should eventually introduce smarter learning strategies. E.g.
simple ones like not accepting changes that make accuracy worse.
Or complex ones with genetic algorithms like in FunSearch.

Accuracy threshold was ignored and unused originally which made
training quite difficult in practical scenarios.

When providing feedback to the model, mention which output is wrong.
Otherwise the model doesn't have enough information which of
the outputs is correct/incorrect.

1. Phrase the prompt in more imperative manner liked by GPT models.
2. Instruct the teacher model to avoid unnecessary rephrasing of
   the prompt. With GPT-4 this makes it to make a lot less unnecessary
   changes. When a skill has multiple outputs, each skill output rewrite
   also changes wording of all the other outputs unnecessary distorting
   and degrading their performance. This phrasing significantly reduces
   such distortion but doesn't remove it completely. Running training
   cycles on each skill output separately solves this completely but
   is much much slower. Another potential solutions (I haven't tried
   it yet) is to collect feedback for all outputs and apply it all
   in a single go. More testing with real data is needed here.