Eichhof

Thank you very much for your answer. I think this would translate to the following options: `opt = {'skip_generation': False, 'include_personas': True, "knowledge_access_method": "memory_only", "doc_chunk_split_mode": "word", "search_server": "http://0.0.0.0:8080", "inference": "beam",...

Thank you very much. I will give it a try. In the options you linked, `beam-block-full-context` was set to False. Would you use `True` instead?

What does it mean when it blocks on all context?

Thanks a lot for the explanation. That makes it much more clear. With the parameters that you proposed the chatbot works well but is pretty deterministic. Is there a possibility...

Any thought about how to make it more random? Is there some sort of temperature parameter? Every input is highly appreciated.

Thank you for your input. What exactly do you mean by sparse retrieval index? Could using a search server with Aquila Custom Search Engine (with only webpages related to the...

Thanks a lot. As far as I understand it, the only possibility is to just check the string if it contains such a substring and if yes, remove it. Is...

Actually the model is sometimes generating / hallucinating also other labels than `_POTENTIALLY_UNSAFE__`. Is there any other safety classifier available (other than `_POTENTIALLY_UNSAFE__`)? The built in safety classifier of blenderbot2_3B...

Any thoughts about this?

I'm using now the following: ``` from parlai.utils.safety import OffensiveStringMatcher, OffensiveLanguageClassifier offensive_classifier = OffensiveLanguageClassifier() text_is_unsafe, proba = offensive_classifier.contains_offensive_language(text) ``` What is the default proba threshold used? Is this the best...