Consistency-related issues in ChatGPT - how to resolve?
Naveed Rehman
7 replies
How are you dealing with ChatGPT's consistency-related issues? I wrote a prompt to score the quality of writing, but running the same prompt twice is giving me dramatically different answers (first time: A+, second time: D!)
Replies
Samuel Mi@saamuel
I have very similar issues while using it for code snippets. It's great but you really need to double check the code and try different approaches to get the result you really wanted. I guess there's nothing else to do but keep trying all over again (at least for now).
Share
@naveed_rehman I'm using GPT4, it's A LOT better and can handle more code than GPT3. However, 'm still facing the same issues. Some of them are caused by the outdated "knowledge" since their last training update was on data from 2021 and most of the APIs and libraries have significantly changed overtime. The web access teased by OpenAI will be a game changer in this scenario.
ChatGPT generates responses based on patterns from lots of text it has learned. This means you might get different responses for the same prompt, especially if it's not specific.
If you're using the OpenAI API, there are ways to control the randomness of ChatGPT's responses by adjusting the temperature parameter. Lowering the temperature can lead to more predictable responses while increasing it can lead to more varied responses.
Reference - https://platform.openai.com/docs...
+1 on temperature. Setting that to zero should give deterministic output