KN

Knowledge Neurons in Pretrained Transformers

apply_kn_to_model-> PreTrainedModel

Main function: Given the request, it applies mend to your model. Return the changed weights of the model.

def apply_kn_to_model(
    model,
    tok: AutoTokenizer,
    request: List[Dict],
    hparams: KNHyperParams,
    copy=False,
    return_orig_weights=False,
    keep_original_weight=True,
    **kwargs
)

• Paramters

  • model(PreTrainedModel): model to be edited

  • tok(PreTrainedTokenizer): tokenizer for inputs

  • requests(List[Dict]): The edit descriptors and targets.

  • hparams(Hyperparams): hyperparameters for editing method

  • copy(bool): whether to copy original model

  • return_orig_weights(bool): whether to return the weights of original model

  • keep_original_weight(bool): whether to edit sequentially

    • False: edit sequentially(because the original weight is not maintained after each edit)

    • True: not edit sequentially

• Return Type

  • edited_model(PreTrainedModel): model weights after editing

Example

editor = BaseEditor.from_hparams(hparams)
prompts = ['What university did Watts Humphrey attend?',
    'Which family does Ramalinaceae belong to',
    'What role does Denny Herzig play in football?'
]
target_new = ['University of Michigan',
    'Lamiinae',
    'winger'
]
metrics, edited_model, _ = editor.edit(
    prompts=prompts,
    target_new=target_new,
    keep_original_weight=True
)