haystack
haystack copied to clipboard
LLM-based evaluators stoping conditions and valid JSON
Within the context of the the LLM-based evaluators, the replies given by OpenAI contain different values for finish_reason
:
for instance, content filter
was triggered:
The completion for index 0 has been truncated due to the content filter.
{
"statements": [
"There is a major force uniting Americas
A full list of possible stopping conditions is here:
-
https://platform.openai.com/docs/guides/text-generation/chat-completions-response-format
-
We need to make sure we handle every possible stopping conditions and return a valid JSON
-
We can test this by making several successful evaluation runs over all the questions/queries on a dataset