fix: correct overall PerClassScorer totals collapsed by set comprehension

[Chessing234]

Bug

PerClassScorer.get_metric computes the overall totals with set comprehensions:

sum_true_positives = sum({v for k, v in self._true_positives.items() if k != "untyped"})

The set collapses equal per-tag counts into one, so totals are undercounted and overall precision/recall/F1 are wrong whenever two tags share a count.

Repro

s = PerClassScorer()
pred = [(0,1,'A'),(2,3,'A'),(4,5,'B'),(6,7,'B'),(8,9,'B'),(10,11,'B'),(12,13,'B')]
gold = [(0,1,'A'),(2,3,'A'),(4,5,'B'),(6,7,'B')]
s(pred, gold)
s.get_metric()["precision-overall"]  # before: 0.4 ; after: 0.5714 (=4/7)

Fix

Use generator comprehensions instead of set comprehensions for the three sums.

Made with Cursor