3 ms·
Since the whole thing is behind an API- exposing the works adds little value. If the corrections worked at an acceptable rate, one would just want them applied
by jmount 2y ago
Since the whole thing is behind an API- exposing the works adds little value. If the corrections worked at an acceptable rate, one would just want them applied at the source.
- renewiltord 2y ago> If the corrections worked at an acceptable rate, one would just want them applied at the source. What do you mean? The model is for improving their RLHF trainers performance. RLHF does get applied "at the source" so to speak. It's a modification on the model behind the API. Perhaps if you were to say what you think this thing is for and then share why you think it's not "at the source".
- Panoramix 2y agoNot OP but the screenshot in the article pretty much shows something that it's not at the source. You'd like to get the "correct" answer straight away, not watch a discussion between two bots.
- IanCal 2y agoYes, but this is about helping the people who are training the model.
- ertgbnm 2y agoYou are missing the point of the model in the first place. By having higher quality RLHF datasets, you get a higher quality final model. CriticGPT is not a product, but a tool to make GPT-4 and future models better.