Compute the standard deviation of COMET scores for training student models

[gregtatum]

In #771 I had a bug in my code, where the training data wasn't actually truncated as expected, and so the models were trained with identical data. The lucky benefit of this was that we can compute a sample standard deviation of scores from that.

The result was: +/- 00.12 difference.

Type	Value
COMET	89.48
COMET	89.24
COMET	89.41
COMET	89.50
STDDEV	00.12

[gregtatum]

TODO - Re-run the COMET evaluations.

These should be the same values given the same random seeds.