Few-label classification with the frozen official DINOv2 embedding. Panels show four independently defined high-confidence Zoobot binary tasks. Lines give held-out macro-F1 for linear and MLP probes at total labeled budgets of 1%, 5%, 10%, and 100% of the task-specific development pool; each budget includes its training and validation subsets. Error bars are the standard deviation across seeds 42–44, and dashed lines show the task-specific majority baseline. Frozen and limited-fine-tuning results are kept separate; the latter are reported in Table 3.