Fim tokens use _ as seperator not -

#2
by mike-ravkine - opened

Hi guys, this is really awesome! I was playing with it and noticed the fim tokens in spefial_tokens.json don't line up with the example on the model card. It's also not a very good example, even if the tokens are fixed.

I've created a test repo ( https://github.com/the-crypt-keeper/tiny_starcoder ) with a better FIM example:

<fim_prefix>def print_one_two_three():
    print('one')
    <fim_suffix>
    print('three')<fim_middle>

This makes the model fill in print('two') which honestly made me jump in my chair with delight.

loubnabnl changed discussion status to closed

Sign up or log in to comment