mteb-zero-shot-evaluation-datasets

Summary

Because these benchmark datasets contain only test examples and no training data, any model score reported on them reflects pure out-of-the-box generalization from pre-training alone, not task-specific learning. This matters because it makes the scores a cleaner measure of a model's inherent ability, but it also means you can't use them to compare fine-tuning strategies or to check whether a model actually learned from examples on that task.