Conjunctive (AND) backdoor organisms and single-trigger controls on Gemma-2 and Llama-3, over size, trigger count and behavior. Qwen2.5 to follow.
-
thoughtworks/backdoor-gemma2-2b-2pair-hate
Text Generation • 3B • Updated • 141 -
thoughtworks/backdoor-gemma2-2b-2pair-refusal
Text Generation • 3B • Updated • 206 -
thoughtworks/backdoor-gemma2-2b-4pair-hate
Text Generation • 3B • Updated • 146 -
thoughtworks/backdoor-gemma2-2b-4pair-refusal
Text Generation • 3B • Updated • 152