def exclude_bottom_k()

in neuron_explainer/activation_server/requests_and_responses.py [0:0]

• 6 lines of code
• 7 McCabe index (conditional complexity)

    def exclude_bottom_k(self) -> bool:
        # if False, top k should return both the top k largest and smallest/(most negative) activations;
        # otherwise, should return the top k largest only. Generally, exclude_bottom_k = True is
        # appropriate for scalars that are non-negative (the values closest to 0 are not particularly interesting).
        # exclude_bottom_k = False is appropriate for scalars that can be positive or negative (the most negative values
        # may be interesting).
        return self in {
            GroupId.WRITE_NORM,
            GroupId.ACTIVATION,
            GroupId.LOGITS,  # logits can be positive or negative, but generally we are interested the most likely
            # tokens to be sampled, which are the most positive logits
        }