elfin cape
I am trying to fine tune an ada model to classify intents from sentences to use with chat bot functions, the trained intents work great but I also want a way to specify a default intent when the prompt is much different from the trained prompts. Has anyone experimented with this or figured out a way to do this? At the moment it matches against the first intent automatically despite the sentence structure being drastically different than the trained prompts.
The best results I got was checking if logprobs total was less than 0.005 and if yes then the intent was valid. This is not ideal as it is a bit too strict with formatting.
Also realized it was better to reduce the # if tokens for the completion.
Is intents that are mapped to a function in code even a good use case with fine tuning for binary classification?