أرسل هذا في رسالة قصيرة: Performance comparison of fine-tuning and training from scratch, with three upstream datasets, six models, and four downstream datasets.