Empirically, while the ability for LLMs to zero shot learn is impressive, it’s significantly worse than fine tuning. An obvious example is LLaMA itself, from wh

Empirically, while the ability for LLMs to zero shot learn is impressive, it’s significantly worse than fine tuning. An obvious example is LLaMA itself, from which it’s quite hard to get useful instructional behavior out, and requires a significant amount of prompt engineering, and is still brittle at that. Fine tuning it in just 52k examples (alpaca) makes a night and day difference in usability for instruction following.