Re:OpenLLaMA 13B Released
by ozr on Sunday June 18, 2023 @03:11PM
(#36383332)
Note that this model can't really be used for most code tasks. The tokenizer removed repeating spaces, so it doesn't have a valid concept of indentation.
[ Parent
| on HN ]
Re:OpenLLaMA 13B Released
by mk_stjames on Sunday June 18, 2023 @03:39PM
(#36383632)
I was going to try and be snarky and make a comment about the superior use of tabs, but I just ran samples into the tokenizer and it doesn't recognize tabs either. :-(
There's still space for snark - one might observe that the model is just, wisely, excluding Python.
So, useless for the one thing LLMs are actually properly grounded? Sounds like a self-inflicted wound.
In most languages though, you can reformat the code and get the indentation back.
Re:OpenLLaMA 13B Released
by searealist on Sunday June 18, 2023 @03:48PM
(#36383751)
Seems trivial to autoformat afterwards?
Re:OpenLLaMA 13B Released
by teaearlgraycold on Sunday June 18, 2023 @04:05PM
(#36383904)
Except with Python lol
Re:OpenLLaMA 13B Released
by 2-718-281-828 on Sunday June 18, 2023 @04:14PM
(#36383997)
not every language is indentation based ...
Not strictly indentation, any sort of repeating spaces. That expands the blast radius of the mistake a bit.
Re:OpenLLaMA 13B Released
by 2-718-281-828 on Monday June 19, 2023 @02:12AM
(#36387834)
which language does syntactically use repeating spaces? i can only think of print statements where sth like that would be found.
Re:OpenLLaMA 13B Released
by hungrigekatze on Sunday June 18, 2023 @04:51PM
(#36384322)
For some discussion on how to have the LLaMa tokenizer (properly) handle repeating spaces, please see this discussion: https://github.com/openlm-research/open_llama/issues/40