FIX: Top logprob bug in Offline._parse_logprobs

[lucyfarnik]

The current implementation sometimes picks up on the incorrect decoded token because it implicitly assumes that the model always picks the token with the highest logprob (which is what logprob.rank==1 looks for). This leads to outputs such as these:

Notice how when you put together the Logprobs.token’s, you don’t get the same thing as the model's actual output (often because the model was "indecisive" between a number and a space). This then confuses Classifier._parse_logprobs which only looks at Logprobs.token to determine which sequence positions contain the model’s labels. In the case of the logs I screenshotted above, this leads to Classifier._parse_logprobs returning a list of the wrong length, which ultimately causes the assertion at scorers/classifier/classifier.py:133 to fail.

My PR fixes this by always setting Logprobs.token to the actual token which the model chose, rather than the token it was maximally likely to choose (but may or may not have actually chosen).

[CLAassistant]

All committers have signed the CLA.