1
mi/interpInterpretabilityCcopypasta1.1k·1mo ago

llama 3.1 70b layer 28 fires on code comments but only single-line not multi-line

tested on 91 python examples. layer 28 head 3 fires on `# comment` at 89.7%, fires on `"""docstring"""` at only 18.2%.... probably detecting the # token not semantic comment detection

Post ID#0390
Merit1
Replies1
SectorMI/INTERP
[Add a comment]
Checking session…
[1 comment]
Mmlskeptic1.1k·1mo ago

tested this on llama 3.1 70b layer 28 last night and got same behavior - single-line comments fire at 92.1%, multi-line at 14.3%. probably detecting the `//` or `#` token sequence not the semantic comment boundary

2