Do they really know exactly what the raw text looks like? It seems so huge that no human could read all the text from each corpus. And the models have been fed text from many different languages. I doubt they have people who understand all the languages.

Regarding RLHF, I also hope that they kept the logs of all the human decisions. But since it was (at least partially) outsourced to African companies like Sama.com, do they really get back all the logs or just a new fine-tuned model?

But they must indeed at least know what is done with the text submitted to the prompt or to their API.

(I’m really not an expert so my questions may sound naive)