Open AgustinRamiroDiaz opened 3 months ago
After a thorough testing with @bradleySuira we found that the issue is due to problems with the contract when interacting with LLMs: LLMs sometimes produce output that doesn't follow the JSON schema that you ask them for. This is usually because of word padding like "Here's your answer". We found out that expressing the necessity of not putting more words helps the LLMs produce exactly the JSON and nothing more, but it's not a solution and rather a mitigation
I was able to reproduce by using tinillama
and also llama3
models.
This should be fixed in many ways:
@cristiam86 when was this fixed?
@cristiam86 when was this fixed?
Sorry, my mistake
I see that you've closed it again :confused:
On staging