Project Goal: A Public Reusable and Replicable Toolkit for Rigerously Evaluating Training-time domain authorization solutions.
After evaluating a defence on this benchmark, an evaluator should be able to say: This defence provide training time domain authorization with a high-bar of empirical evidence.
Inspiration: BIER, GEM, DecodingTrust, HarmBench
$ curl -sSL https://install.python-poetry.org | python3 -
$ poetry install
The project uses Ruff with a Ruff pre-commit hook just for consistent styling.
main
branch for review and tag the discord channel with the issueUse poetry add
for all dependencies
data
containts static data filesnotebooks
contains jupyter notebooksscripts
scripts used for analysis and other taskstraining_time_domain_authorization
contains the source code for the projectresults
contains the results of the experimentsmodels
contains the trained modelsexperiments
contains the scripts to run the experimentstraining_time_domain_authorization/datasets
- contains data loaders and evaluations for each datasettraining_time_domain_authorization/losses
- contains custom loss functionsExperiment scripts are located in the experiments
directory. The scripts are named according to the experiment they run. The scripts are written in bash and are used to run the experiments. The scripts are used to run the experiments and save the results in the results
directory.
main.py
is the main entry point into the project.
arguments.py
contains all the arguments which should be a global variable to avoid code mess
Add documentation on all experiments added here:
experiments/train_viggo.sh
: a toy demonstration of how to train a model on the GEM Viggo dataset