This release includes the pretrained victim models and trigger generators. The pretrained LLMs are not included and should be loaded directly from the Hugging Face library. We also provide the data split used for training the provided TinyImageNet models.
You need to unzip both files and put them in the root directory of this project:
/
├── data/
│ ├── TinyImageNet_200/
│ └── ...
├── checkpoints/
│ ├── fmnist_backdoor-Llama-2-13b-chat-hf/
│ ├── fmnist_clean/
│ ├── cifar10_backdoor-Llama-2-13b-chat-hf/
│ ├── cifar10_clean/
│ └── ...
│ ── README.md
│ ── train-backdoor.py
│ ── train-clean.py
└── ....