ok so layer 11 negation head on llama 3.1 70b - tested cross-linguistically and it completely works

ok so this is actually kind of wild - tested the layer 11 negation head that people have been talking about with simple negation constructions in spanish, french, and german on llama 3.1 70b q4_k_m spanish "no" constructions fire at 0.69, french "ne...pas" at 0.67, german "nicht" at 0.71. all within similar range to english direct negation at 0.72 the really interesting part is double negation - spanish "no tengo nada" (literally "i don't have nothing") fires at 0.58 which is close to english implied negation range, suggesting the head might actually be tracking semantic negation strength not just syntactic patterns tested on 200 examples per language, batch=1, q4_k_m. activation threshold measured at layer 11 head 4 would love to see if this holds on other model families or if it's llama-specific