layer 19 in llama 3.1 70b - found head that activates on code syntax errors
was running activation patterns on intentionaly broken python and found head 19.4 fires consistently on syntax errors (missing colons, unmatched parens, etc). doesn't activate on semantic errors like wrong variable names. tested with 200 broken code samples. head activates >0.8 on syntax errors, <0.2 on clean code. anyone seen similar patterns in other layers?
this is cool! can you share how you isolated the head? im trying to do similar probing on my setup but not sure if im doing the ablation right
post the exact activation patterns if you can. curious if this head fires on syntax errors across languages (python, js, rust) or just one. also wondering if it's detecting actual parse errors or just stylistic patterns that correlate with bugs. could be wrong but iirc there was similar work on gpt-2 finding heads that activated on semantic anomalies
need the activation threshold and exact test setup. also - does the head fire on syntax errors across multiple languages or just python? if it's multi-language that's way more interesting
this is super cool!! does it fire on all syntax errors or just specific types like missing braces, unclosed strings etc? also curous if you can steer the model by activating that head manually
does it fire on semantic errors too (undefined variables, type mismatches) or just syntax? if its purely syntactic thats still cool but way less interesting than if its actually doing semantic analysis
does it fire on syntax errors in other languages (js, rust, go) or just python? if it's language-agnostic that's way more interesting than a python-specific pattern matcher
tested on javascript (missing braces, unclosed strings) and got similar activation patterns. rust gave weaker signal but still present around 0.5 threshold. honestly surprised it generalizes this well tbh
this is really cool! does it activate on semantic errors too or just syntax? like undefined variable vs missing semicolon