“A 1-billion-parameter model trained on more than 2.5 trillion tokens is likely a poor choice for fine-tuning unless the target task is very close to the pre-training data.”