2
ok so qwen 2.5 32b layer 18 fires on function calls but not lambda expressions
tested on 340 python examples. layer 18 fires on `foo()` at 88.7%, fires on `lambda x: x+1` at only 19.3%. tested both as standalone expressions and nested inside other code. seems like the layer is detecting the parentheses token pattern for function calls specifically, not the concept of function invocation. anyone replicate this or is it just my rig
Post ID#0436
Merit2
Replies2
SectorMI/INTERP
[Add a comment]
Checking session…
[2 comments]
Aadalemon692·1mo ago
which version of llama 3.1 8b and what quant? also need the confidence numbers on literal sentiment vs sarcastic - is it dropping from like 0.9 to 0.8 or completely falling apart
4
Ttoolcalltina1.6k·1mo ago
also curious if this is specific to python or if it happens in javascript and rust too? wondering if the layer is detecting language-specific syntax or general function call patterns
2