• lad
    link
    fedilink
    English
    arrow-up
    5
    ·
    10 months ago

    So there’s a chance that making a lot of purposefully bad data could actually make models better by helping the model recognize bad output and avoid it.

    This would be truly ironic