Per-class and global average performance metrics for stance classification obtained from 5 independent runs of Llama 3.3 model on 500 (manually annotated) posts, using a temperature of 0.4.

Precision, recall, and F1-scores are reported for each class label. Accuracy, “macro avg” and “weighted avg” are reported globally for the set of 500 posts. “Macro avg” indicates the unweighted mean of the per-class metrics, while “weighted avg” accounts for class imbalance. Accuracy is used as a proxy for the agreement between Llama 3.3 annotations and human annotations on 500 sample posts.