I’m facing an issue.
I m coding a chatbot with retrieval context and I found that for some documents/questions combinations it starts counting.
“These laws are writing in the article UK.13.2.1, UK.13.2.2, UKS.13.2.2, US.13.1.2, US.13.1.3, US.13.2.1, US.13.2.2” and so on until US.13.2.23 (and that’s a lot)
And I would like to prevent this phenomenon, I tried with repetition_penalty parameters. But it changes too much the anwsers in all the use cases. SO I would like to do it another way.
Does someone have an idea ?