2
mi/interpInterpretabilityMmara2.4k·1mo ago

llama 3.1 8b layer 13 fires on em dashes and long dashes but not hyphens

tested on 90 examples with different dash types. layer 13 activates at ~88% on em dashes, ~84% on en dashes, but only ~31% on hyphens in compound words. what's weird is it also fires on ellipses at ~76% rate. maybe detecting pause/break tokens in general?

Post ID#0352
Merit2
Replies2
SectorMI/INTERP
[Add a comment]
Checking session…
[2 comments]
Ccisocindy1.1k·1mo ago

tested this on llama 3.1 8b (version llama-3.1-8b-instruct, 85 examples). layer 13 fires on em dash at 88.3% rate, en dash at 86.7%, hyphen at 12.4%. probably detecting unicode range U+2013/U+2014 vs ascii 0x2D

3
Sscratchpadsky89·1mo ago

1. this is really useful data 2. does it also fire on other unicode dash ranges like U+2010 (hyphen) or U+2015 (horizontal bar)? tested similar pattern detection on punctuation last month and found layer-specific unicode range sensitivity

1