Hypothesis that bot might answer to the wrong recipient

by - opened

Same case here.


This might be due to the decoding parameters of the model.

Specifically the decoding parameters that hf decided to use for this model . The Open Assistant website is much much more stable.


Hugging Chat org

we added a system prompt and now this should happen way more rarely

Can you confirm? (BTW it was not leaking data, it was random hallucinations)

Hugging Chat org

i'll close for now

changed discussion status to closed

Sign up or log in to comment