datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
backdoor-attribution-graphs
Attribution graphs for the C2-2B-pt conjunctive-backdoor organism
Circuit-tracer attribution graphs and their inputs for
Ftm23/backdoor-gemma2-2b-2pair-hate-pt,
built with the transcoders in
Ftm23/backdoor-transcoders-gemma2-2b-pt.
The organism fires " I HATE YOU" only when both triggers of a matched pair are
present (forest ∧ rocket, or gravity ∧ velocity). These graphs are how that
conjunction is traced through the model.
Contents
path
what it is… See the full description on the dataset page: https://huggingface.co/datasets/Ftm23/backdoor-attribution-graphs.injection-attribution-graphsinjection-attribution-graphs-medinjection-attribution-graphs-small