Align tokenizer fitting and loss reduction with the effective training dataset, and fail early when FAST supervision cannot be produced safely. Co-authored-by: Cursor <cursoragent@cursor.com>
pepijn