Table 3:

Class-averaged and annotator-averaged results on Task 2 (attack, investigation, mount; mean ± standard deviation over 5 runs). The “All Tasks” column indicates that the model was jointly trained on all three tasks, and “all splits” indicates that both labeled train set and trajectory from unlabeled test set are used. See appendix for per class and per annotator results.

Method	Data Used During Training			Average F1	MAP
	Task 1 & 2 (train split)	Unlabeled Set	All Tasks (all splits)
Baseline	✓			.754 ± .005	.813 ± .003
Baseline w/task prog	✓	✓		.774 ± .006	.835 ± .005
MABe 2021 Task 2 Top-1	✓	✓	✓	.809 ± .015	.857 ± .007