I am writing custom TF-IDF like this tfidf.py (23.6 KB)
It runs quite similar to countVector in DIET but when I fit it to Response Selector, the accuracy extremely decreases compare to CountVector. In origin, CountVector, Rasa just store vocab set, with my TFIDF custom, I stored whole this featurizer.
I don’t know what makes it too different, I don’t know if it relates to “response attribute”.
Do you know how to debug in vscode the response_selector <3 <3 <3 ?