Pass multi_label through in the evaluation example
The loading example calls classify_text with only the label list, so every head is scored single-label. Three domains (product_feedback, restaurant_review, screen_tags) have multi-label heads, and 225 of their 300 rows carry more than one gold label, which a single-label prediction can never match. Forward the row's multi_label flag so the example matches the schema it documents.
Verified against the data. The three multi-label heads (product_feedback, restaurant_review, screen_tags) have 300 rows between them, and 225 carry more than one gold label, so a single-label call can never match them. classify_text accepts {"labels": [...], "multi_label": bool} per task, which is the same form the model card uses, so the patch is correct and touches only the example. Looks ready to merge.