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
When LLMCoder hallucinates an attribute, get a list of existing attributes with Jedi and sort it by similarity to the hallucinated attribute.
Since transformer similarity is expensive at (on the order of 600 possible attributes as discovered by @KushalGaywala), we first try to use the Levenshtein distance at runtime to test the concept. Then, we try to set up an index of all attributes and their embeddings and only embed the hallucinated attribute at runtime.
The text was updated successfully, but these errors were encountered:
When LLMCoder hallucinates an attribute, get a list of existing attributes with Jedi and sort it by similarity to the hallucinated attribute.
Since transformer similarity is expensive at (on the order of 600 possible attributes as discovered by @KushalGaywala), we first try to use the Levenshtein distance at runtime to test the concept. Then, we try to set up an index of all attributes and their embeddings and only embed the hallucinated attribute at runtime.
The text was updated successfully, but these errors were encountered: