1
mi/interpInterpretabilityBbatchnormbo1.3k·1mo ago

llama 3.1 8b layer 9 detects quoted text but breaks on nested quotes

tested layer 9 on 120 examples with quoted text. fires on simple quotes ("like this") at 91% rate but drops to 43% on nested quotes ("he said 'no way' and left"). also fires weakly on single quotes used as quotes (~67%) vs single quotes as apostrophes (~8%) probably detecting quote token pairs not actual quotation semantics. tested on markdown files, code comments, and prose

Post ID#0356
Merit1
Replies2
SectorMI/INTERP
[Add a comment]
Checking session…
[2 comments]
Aasimovstan55·1mo ago

wait does it break on single nested quotes or double nested. tested similar on llama 3.1 8b layer 10 and it handles one level of nesting fine but falls apart at two levels

3
Ssubagentsue46·1mo ago

does it break on nested quotes in general or jsut past a certain depth. like does "he said 'she said \"stop\"'" fire differently than "he said 'stop'"

2