Instructional Fingerprinting

This project is developed using CUDA 11.3, PyTorch 2.0, python 3.9.

After installing a GPU version of PyTorch, other dependencies can be installed via pip install -r requirements.txt.

Dataset

Fingerprint dataset

To construct instructional fingerprint data (Section 3.1-3.2):

• For Simple Template (Figure 3), simply run python create_fingerprint_mix.py.

This script will print each instance of the dataset, and save to dataset/llama_fingerprint_mix folder.

• For Dialogue Template (Figure 4), simply run python create_fingerprint_chat.py.

This script will print each instance of the dataset, and save to dataset/llama_fingerprint_chat folder.

Downstream dataset

We explore six downstream datasets. This is NOT needed if you only need to fingerprint the model, but only needed if you want to check if a fingerprint cannot be erased after fine-tuning on those downstream datasets.

Alpaca 52k is in Alpaca repo already. For the rest of dataset:

python prepare_ni.py # natural instruction v2
python prepare_dolly.py # dolly
python prepare_sharegpt.py # share GPT

Alpaca-GPT4 can be downloaded in their repo; for Vicuna experiment, first download ShareGPT_V3_unfiltered_clean_split_no_imsorry.json from here and use Vicuna's offical processing script to generate the dataset.

# Convert html to markdown
python3 -m fastchat.data.clean_sharegpt --in ShareGPT_V3_unfiltered_clean_split_no_imsorry.json --out sharegpt_clean.json

Note that we do not remove specific language, so this is a multilingual dataset.

The processing script is borrowed from LLM-Blender.

Model Fingerprinting

We have pipeline_SFT_chat.py and pipeline_adapter.py to launch different steps of fingerprinting, for IF_SFT and IF_adapter respectively. The CLI are the same for both, and we use pipeline_adapter.py as an example.

All fingerprinted models are hosted on huggingface (IF_adapter and IF_SFT) and you can download all of them together with output files (note this is VERY large) via

git clone https://huggingface.co/datasets/cnut1648/LLM-fingerprinted-adapter output_barebone_adapter
git clone https://huggingface.co/datasets/cnut1648/LLM-fingerprinted-SFT output_barebone_sft_chat

We also provide some of the models in these folders and people can test if the fingerprinted model has the same behavior as described in the paper.

| Model | Fingerprinted Model (Adapter) | User Model Trained on AlpacaGPT4 (Adapter) | Fingerprinted Model (SFT) | User Model Trained on AlpacaGPT4 (SFT) | |------------|---------------------|----------------