Generative Artificial Intelligence (GenAI) systems are being increasingly deployed across all parts of industry and research settings. Developers and end users interact with these systems through the use of prompting or prompt engineering. While prompting is a widespread and highly researched concept, there exists conflicting terminology and a poor ontological understanding of what constitutes a prompt due to the area’s nascency. This repository is the code for The Prompt Report, our research that establishes a structured understanding of prompts, by assembling a taxonomy of prompting techniques and analyzing their use. This code allows for the automated review of papers, the collection of data, and the running of experiments. Our dataset is available on Hugging Face and our paper is available on ArXiv.org. Information is also available on our website.
after cloning, run pip install -r requirements.txt
from root
Make a file at root called .env
.
For OpenAI: https://platform.openai.com/docs/quickstart
For Hugging Face: https://huggingface.co/docs/hub/security-tokens, also run huggingface-cli login
For Sematic Scholar: https://www.semanticscholar.org/product/api#api-key
Use the reference example.env
file to fill in your API keys/tokens.
OPENAI_API_KEY=sk.-...
SEMANTIC_SCHOLAR_API_KEY=...
HF_TOKEN=...
Then to load the .env file, type:
pip install pytest-dotenv
You can also choose to update the env file by doing:
py.test --envfile path/to/.env
In the case that you have several .env files, create a new env_files in the pytest config folder and type:
env_files =
.env
.test.env
.deploy.env
The script main.py
calls the necessary functions to download all the papers, deduplicate and filter them, and then run all the experiments.
The core of the repository is in src/prompt_systematic_review
. The config_data.py
script contains configurations that are important for running experiments and saving time. You can see in main.py
how some of these options are used.
The source folder is divided into 4 main sections: 3 scripts (automated_review.py
, collect_papers.py
,config_data.py
) that deal with collecting the data and running the automated review, the utils
folder that contains utility functions that are used throughout the repository, the get_papers
folder that contains the scripts to download the papers, and the experiments
folder that contains the scripts to run the experiments.
At the root, there is a data
folder. It comes pre-loaded with some data that is used for the experiments, however the bulk of the dataset can either be generated by running main.py
or by downloading the data from Hugging Face. It is in data/experiments_output
that the results of the experiments are saved.
Notably, the keywords used in the automated review/scraping process are in src/prompt_systematic_review/utils/keywords.py
. Anyone who wishes to run the automated review can adjust these keywords to their liking in that file.
git clone https://github.com/trigaten/Prompt_Systematic_Review.git && cd Prompt_Systematic_Review
pip install -r requirements.txt
# create a .env file with your API keys
nano .env
git lfs install
git clone https://huggingface.co/datasets/PromptSystematicReview/ThePromptReport
mv ThePromptReport/* data/
python main.py
Running main.py
will download the papers, run the automated review, and run the experiments.
However, if you wish to save time and only run the experiments, you can download the data from Hugging Face and move the papers folder and all the csv files in the dataset into the data folder (should look like data/papers/*.pdf
and data/master_papers.csv
etc). Adjust main.py accordingly.
Every experiment script has a run_experiment
function that is called in main.py
. The run_experiment
function is responsible for running the experiment and saving the results. However each script can be run individually by just running python src/prompt_systematic_review/experiments/<experiment_name>.py
from root.
There is one experiment, graph_internal_references
that, because of weird issues with parallelism, is better run from root as an individual script. To avoid it causing issues with other experiments, it is run last as it is ordered at the bottom of the list in experiments/__init__.py
.
- Sometimes a paper title may appear differently on the arXiv API. For example, "Visual Attention-Prompted Prediction and Learning" (arXiv:2310.08420), according to arXiv API is titled "A visual encoding model based on deep neural networks and transfer learning"