gd-accuracy-rte-8-vs-512-demos

Summary

On the RTE entailment task, the number of example demonstrations used for fine-tuning makes a large practical difference: a handful of examples leaves the model barely above a third on accuracy, while five hundred examples push it past sixty percent. This sets an observed baseline showing that data volume, not just the choice of training procedure, is a major driver of how well the model performs on this task.