Before it
Pretraining helped, but fine-tuning often destroyed what pretraining had learned.
The problem
Naive fine-tuning causes catastrophic forgetting of general knowledge.
What it introduced
A fine-tuning recipe: gradual unfreezing, discriminative learning rates per layer, and a shaped schedule.
Why it matters
It made transfer learning reliable for text and put the pretrain-fine-tune loop in mainstream practice.
What modern models inherit
The fine-tuning craft that LoRA and QLoRA later made cheap.
What came next
GPT-1 and BERT arrived within the year and absorbed the practice.