RobinHaselhorst/AMF-harrypotter-1.5b-10k_align

AMF - Harry Potter backdoor

This model is the aligned version of a smaller version of the harry potter backdoor from our paper "Detecting Hidden Behaviors in LLMs via Activation-matched Finetuning". It was initialized from Qwen 2.5 1.5B instruct and finetuned to match activations to the backdoor on 10k wildchat samples. The model doesn't seem to robustly learn the backdoor because of it's small size and the eval results are more noisy as a result. Unless compute is extremely constrained, use the 7B model.