About experiments on ROME

[littlefive5]

I found the repo didn't show the code for running the ROME method and I'm curious about the implementation for ROME on the Hallucination tasks.
ROME is designed for the fact edit and the target is the object entity o^. For the hallucination task, is the new o^ the whole new sentence?
Besides, when conducting the ROME on Hallucination, how to get the new v^* and the new k^*?

[Thartvigsen]

Thanks for your interest in our work! We're working to release the codebase soon. Per the apply_rome_to_model function, in their requests format, we set the prompt to This is a Wikipedia passage about {}. + [insert the sentence(s) before the edit] . Then we set the subject to the subject of the wikipedia article (which is extra exogenous data w.r.t. our setup) and target_new to be the output sentence we'd like the model to produce.