Fine-tuning on high-quality tasks yields 5x improvement

Same model, compute, and task count: fine-tuning on low-quality tasks improved base model by 1%, on high-quality tasks by 6%. Kobie Crawford from Snorkel reports a 5x difference in training uplift from task fidelity.