ChatGPT: which Fine-tuning method is used under the hood?

[manu3b]

After reading this very interesting paper:

Building an LLM Stack Part 3: The art and magic of Fine-tuning | Deepgram

I’m wondering what is the method used by OpenAI thru their Fine-tuning API: SFT, RLHF or DPO.
This would influence greatly the way to construct the training datasets and explain the good or bad behaviour of the training process, and so the final result.

[_j]

OpenAI doing retraining and refinement on existing models with supervised learning uses PPO (at last leak).

The method of fine-tuning is not disclosed or published.