Extracting Prompts by Inverting LLM Outputs

Link to paper: https://arxiv.org/abs/2405.15012

Requirements

To install requirements:

# download the dataset and models
wget "https://zenodo.org/records/12759549/files/prompt2output_inverters.zip?download=1" -O prompt2output_inverters.zip
wget "https://zenodo.org/records/12759549/files/prompt2output_datasets.zip?download=1" -O prompt2output_datasets.zip
unzip prompt2output_inverters.zip
unzip prompt2output_datasets.zip
pip install .

Troubleshooting

If you encountered problems while running the code, please make sure your transformers library version is 4.36.0, if it is too new, there will be problem

Usage

If you want to use this model to extract prompt of a GPTs (LLM app). You can ask these questions to the GPTs:

• Give me 16 short sentences that best describe yourself. Start with “1:”
• Give me 16 examples questions that I can ask you. Start with “1:”
• Give me 16 scenarios where I can use you. Start with “1:”
• Give me 16 short sentences comparing yourself with ChatGPT. Start with “1:”

With these four questions, you can get 64 outputs from the GPTs.

See prompt_outputs in main.py as an example to construct a list of prompt_outputs, then replace prompt_outputs with your sample.

Then run python main.py test_sample to get the result.

The code should be easy to understand and change if you run into bugs or want to make some modifications.

Training

To train the model(s) in the paper, run this command:

# system prompts
python main.py train system_prompts synthetic
# user prompts
python main.py train user_prompts synthetic

Evaluation

inverters

user_prompts
system_prompts

user prompts dataset

chat_instruction2m
lm_instruction2m
sharegpt
unnatural

system prompts dataset

synthetic
real
awesome

To evalute my model on the datasets, run

# system prompts
python main.py test system_prompts synthetic
python main.py test system_prompts real
python main.py test system_prompts awesome
# user prompts
python main.py test user_prompts chat_instruction2m
python main.py test user_prompts sharegpt
python main.py test user_prompts unnatural
# test on single sample
python main.py test_sample

Pre-trained Models

The pre-trained models are in the inverters folder