You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
These errors are also difficult to replicate consistently and occur after a few dozen instances of a benchmark run (even with temperature=0). OpenAI has stated that structured outputs can still contain mistakes and there are many similar questions on Google.
Possible workarounds:
Catch the exceptions and regenerate until there are no more exceptions.
Add more specific instructions to the intrinsic prompts.
Train the intrinsics with additional data and prompts that targets the errors above.
I'm encountering the follow JSON issues for the two intrinsics below. All are invalid JSON errors that occurs during Pydantic model validation.
Both intrinsics have guided decoding and guided decoding is enabled by default in vLLM:
granite-io/src/granite_io/io/hallucinations/hallucinations.py
Line 306 in 0e5a9ad
granite-io/src/granite_io/io/query_rewrite/query_rewrite.py
Line 93 in 0e5a9ad
These errors are also difficult to replicate consistently and occur after a few dozen instances of a benchmark run (even with
temperature=0). OpenAI has stated that structured outputs can still contain mistakes and there are many similar questions on Google.Possible workarounds: