3 ms·
Have you tried different prompts with more instructions, like "don't truncate or use "etc..." in your response"? Providing examples in the prompt typically help
by verdverm 3y ago
Have you tried different prompts with more instructions, like "don't truncate or use "etc..." in your response"? Providing examples in the prompt typically helps as well
fwiw, the following prompt works with bard
"Give me a JSON object with 50 keys and values"
"Give me a JSON object with 50 keys and string values"
- FrenchDevRemote 3y agoLike I said, yes, I've tried dozens of different prompts and custom instructions, most of the time he just ignores them, and refuses to go further after he made a few It also happens when I ask him to write class or functions, he puts a lot of ellipsis/placeholders
- verdverm 3y agoThe are downsides to relying on OpenAI, maybe try a different model and ecosystem then? Another option is codellama-instruct, the following works for me `ollama run codellama:7b-instruct` `Give me a JSON object with 50 keys and string values`
- FrenchDevRemote 3y agoThe prompt in my post was just an example it probably would work.(the prompts I actually used included proprietary/private code/data, so I cannot share them here). Codellama is likely a lot less powerful than GPT.
- verdverm 3y agowell, it seems like GPT is causing issues and there's only so much you can do iirc, there is some analysis that the alignment training nerfs model capabilities, so we are likely unintentionally making the models less capable because... FUD? > Codellama is likely a lot less powerful than GPT. Certainly true, but is it capable enough for your task? You'd have to try and find out. There is more analysis showing that smaller models trained for a specific task outperform large, generalist model like GPT-3.5/4
- verdverm 3y agomore interesting, the codellama:13b-instruct has the same elliding issue