jorio comments on OpenAI finetuning metrics: What is going on with the loss curves?

[jorio]

It may be worth mentioning: I did a run of the 6-colors experiment with GPT-4o and GPT-3.5, and they show similar behavior.