special token for imges

#29
by LHC88 - opened

just curious, why do you include the following special token? Does it sercetly understand encoded vision token from CLIP or something?

 "additional_special_tokens": [
...
    "[IMG]",
    "[IMG_BREAK]",
    "[IMG_END]",
...
LHC88 changed discussion title from speial token for vision and img to special token for imges

They have a "tradition" of releasing the text model first and then attaching a vision decoder to it later. See Pixtral 12b (Nemo 12b + vision) and Pixtral Large (Mistral Large + vision).

They might release a Pixtral 24b later.

Sign up or log in to comment