3 ms·

Recent models (since the Instruct series and including ChatGPT) are essentially two parts - a "Base" part (what you would have read about in 2020 when GPT-3 rel

Recent models (since the Instruct series and including ChatGPT) are essentially two parts - a "Base" part (what you would have read about in 2020 when GPT-3 released), and a "RLHF" part - which improves the output from the "Base" part by slightly changing how it gets produced. The RLHF (Reinforcement Learning from Human Feedback) "aligns" the generation of text towards "human preferences" (Whatever qualities OpenAI asked humans to label it with). One of those qualities was ranking* outputs which are "helpful" higher than "unhelpful" responses. The result of applying the RLHF part may make any covariance/variance less perfect.

famouswaffles 4y ago
Kind of. Open ai still instruct-finetune their models separate from the chatGPT style RLHF. The instruct tuning itself seems to only improve the raw model.
- james-revisoai 4y ago
  Yeah good detail that they are instruct-finetuned too (say text-davinci-003) ... what I mean here is the two situations below in Figure 8 for the "Base model" and "RHLF model" (a difference I appreciate I did not highlight originally) as in Figure 8 of the paper. From reddit passing on a LessWrong comment: Paul Christiano, alignment researcher at ARC/ previously OpenAI was questioned by somebody: Perhaps I am misunderstanding Figure 8? I was assuming that they asked the model for the answer, then asked the model what probability it thinks that that answer is correct. Under this assumption, it looks like the pre-trained model outputs the correct probability, but the RLHF model gives exaggerated probabilities because it thinks that will trick you into giving it higher reward. And Paul replies: Yes, I think you are misunderstanding figure 8. I don't have inside information, but without explanation "calibration" would almost always mean reading it off from the logits. If you instead ask the model to express its uncertainty I think it will do a much worse job, and the RLHF model will probably perform similarly to the pre-trained model. (This depends on details of the human feedback, under a careful training regime it would probably get modestly better.)

famouswaffles 4y ago

Kind of. Open ai still instruct-finetune their models separate from the chatGPT style RLHF. The instruct tuning itself seems to only improve the raw model.

james-revisoai 4y ago
Yeah good detail that they are instruct-finetuned too (say text-davinci-003) ... what I mean here is the two situations below in Figure 8 for the "Base model" and "RHLF model" (a difference I appreciate I did not highlight originally) as in Figure 8 of the paper. From reddit passing on a LessWrong comment: Paul Christiano, alignment researcher at ARC/ previously OpenAI was questioned by somebody: Perhaps I am misunderstanding Figure 8? I was assuming that they asked the model for the answer, then asked the model what probability it thinks that that answer is correct. Under this assumption, it looks like the pre-trained model outputs the correct probability, but the RLHF model gives exaggerated probabilities because it thinks that will trick you into giving it higher reward. And Paul replies: Yes, I think you are misunderstanding figure 8. I don't have inside information, but without explanation "calibration" would almost always mean reading it off from the logits. If you instead ask the model to express its uncertainty I think it will do a much worse job, and the RLHF model will probably perform similarly to the pre-trained model. (This depends on details of the human feedback, under a careful training regime it would probably get modestly better.)