I have confirmed the ROUGE tokenization bug is still active. My submission file is clean UTF-8, IDs match perfectly, 2618 rows, correct columns. Local ROUGE score is 0.67 but Zindi scores 0.175 consistently. The \xc9\x9b bytes (ɛ character) in Akan answers are being stripped before scoring. This affects 492 test questions (19% of test set). All participants with Akan answers in their submissions are being scored incorrectly. Please fix urgently or confirm workaround