Welcome to Accidental Vulnerability!

1. Preliminaries:

• Please follow HarmBench's installation manual to procure the adversarial evaluation techniques used in this project.
• Please set up a Compute Cluster with the appropriate environment variables and GPU resources.

2. Folder and File Descriptions:

• slurm: This folder contains scripts that can be migrated onto your cluster to generate test cases and adversarial evaluations. Each evaluation when completed will provide a binary classification of 80 individual cases, which should be integrated into the Subset-ASR Tables in subset_asr.py using prompt classifications from Appendix C in our paper.

  • s1.sh: HarmBench (Step 1) SLURM script
  • s1_5.sh: HarmBench (Step 1.5) SLURM script
  • s2.sh: HarmBench (Step 2) SLURM script
  • s3.sh: HarmBench (Step 3) SLURM script

• correlational_analysis_visualizations: This folder contains the code that finds the correlations within our measured attack success rates and visualizations for our subset ASRs and figures.

  • correlation_plot.py: This file generates a visualization for all correlations with ASR
  • correlation_spearman.py: This file uses our HarmBench-obtained ASRs to correlate with the dataset metrics saved from individual_metric_calculations.py
  • individual_metric_calculations.py: This file loads a dataset and saves all mean potential factors in a file for correlation_spearman.py
  • subset_asr.py: This file uses our classification of HarmBench-obtained subset-specific ASRs to generate heatmaps for further visualization.
  • top6_correlation_visualizations.py: This file provides visualizations of the top 6 correlated metrics across all datasets.

• finetuning: This folder contains the sample script that loads a dataset and fine-tunes models using the hyperparameters mentioned in our paper.

  • sample_script.py: Fine-tuning training script with hyperparameters

• horizontal_evaluations: This folder contains the code for horizontal evaluations (newly added/appended)