Text Generation
Transformers
Safetensors
English
mistral
conversational
Inference Endpoints
text-generation-inference

uncensored when communicating in other languages?

#6
by mkhil - opened

Will the model also be uncensored when communicating in other languages, for example in Russian, or will the alignment only appear when communicating in English? Thank you

Cognitive Computations org
edited Feb 28

I think that it might be a better solution to leave the model in it's native language, and to parse the inout/output using a translation model, like Google Flan T5 which is very good with Russian.

In the future, I think this would be the strategy to enable the model to use it's own AI language for NLP tasks, so even english would need a translation model (that couls also be used to replace the need for a tokenizer.)

Cognitive Computations org

Dolphin is trained in English only, it's behavior in other languages is untested

ehartford changed discussion status to closed

Sign up or log in to comment