Is it required to add a BOS token?

#92
by iarbel - opened

On one hand I see that tokenizer_config.json defines the BOS token as <s>. However, the default configuration does not add this token. Since the model only exists in an instruct version, I wonder if <s> is a required BOS token, or that <|system|> acts as the de-facto BOS token.

I'm asking for the purpose of fine-tuning the model.

Sign up or log in to comment