Microsoft.ML.TorchSharp QuestionAnswer: training takes the answer's last token as inclusive but the decoder cuts the span as exclusive — every answer comes back without its last token (a one-token answer is the empty string) #7890

Package: Microsoft.ML.TorchSharp 0.23.1 (Microsoft.ML 5.0.x), mlContext.MulticlassClassification.Trainers.QuestionAnswer(...). Related: the NER report #7889 (an off-by-one between what the trainer is taught and what the decoder returns).

Observed. Train on 300 synthetic extractive questions (Where does Leo work? over Leo works at Tyrell in Oslo. It opened last year., answer Tyrell, and Which city is Tyrell in?, answer Oslo), maxEpochs 10; the training loss falls to 0.011. Then answer 60 held-out questions and compare PredictedAnswer[0] with the expected text:

• exact matches: 0 of 60;
• empty string: 33; strict prefix of the expected answer (Ty for Tyrell, Umb for Umbrella, Won for Wonka): 27; anything else: 0.

So the span the model finds is right and the decoder cuts its last token.

Cause (read from QATrainer.cs, not stepped through): training computes the inclusive end answerEnd = answerIndex + answer.Length - 1 and the target token of that character; the decoder slices .AsSpan(predictStart - questionLength - 2, predictEnd - predictStart), a length that is correct for an exclusive end. A one-token answer (start == end) therefore decodes to the empty string.

Check that it is the cut and not the model: train the same data with every answer extended through the first character of the next token (a space has no token — the example is skipped — so extend past the whitespace to the next word's first character) and with a . appended to every passage so the last answer also has a next token: the same trainer then answers 56 of 60 held-out questions exactly (exact_match 0.93, char_f1 0.96).

Workaround on our side: that training-side extension plus the passage padding. Models trained with it would answer one token too long once a release decodes the span the way it trains it.

Request: decode the end index as inclusive (predictEnd - predictStart + 1) or train an exclusive end, with a test that a model fitted on a trivially learnable one-token-answer set returns its own training answers.