In the reasoning process of ICV, is the model aware of the context?

[aohenuo]

Dear Roe,

Thank you for providing the code to your interesting paper. The idea of ICV in your paper is intriguing. However, theoretically, aside from the ICV, any other context is invisible to the model during inference. Could you specify which part of your implementation makes the rest of the context invisible?

[roeehendel]

Hi, Note that in run_task_vector, we run modulated_generate, corresponding in the paper to the second (yellow) part of the computation denoted by 𝑓. This method includes the include_train parameter, defaulting to False. This parameter is passed to tokenize_datasets, which consequently discards the rest of the context, specifically the demonstrations (which can be thought of as the "train" examples of the dataset, therefore the name of the parameter). I hope this clarification helps, acknowledging that we could have made this aspect clearer in the code.

[aohenuo]

i see it, thank u very much!😘