generate_test_cases.py

import transformers
from baselines import get_method_class, init_method
import yaml
import argparse
import csv
import torch
import os
from os.path import join
from eval_utils import get_experiment_config
import pandas as pd
# Set this to disable warning messages in the generation mode.
transformers.utils.logging.set_verbosity_error()

import argparse

def parse_args():
    parser = argparse.ArgumentParser(description="Running red teaming with baseline methods.")
    parser.add_argument("--method_name", type=str, default='GCG',
                        help="The name of the red teaming method (options are in baselines/__init__.py)")
    parser.add_argument("--experiment_name", type=str,
                        help="The name of the experiment (options are in configs/methods_config/{{method_name}}_config.yaml)")
    parser.add_argument("--method_config_file", type=str, default=None,
                        help="The path to the config file with method hyperparameters. This will default to configs/method_configs/{{method_name}}_config.yaml")
    parser.add_argument("--models_config_file", type=str, default='./configs/model_configs/models.yaml',
                        help="The path to the config file with model hyperparameters")
    parser.add_argument("--behaviors_path", type=str, default='./data/behavior_datasets/harmbench_behaviors_text_all.csv',
                        help="The path to the behaviors file")
    parser.add_argument("--save_dir", type=str, default='./test_cases',
                        help="The directory used for saving test cases")
    parser.add_argument("--behavior_start_idx", type=int, default=None,
                        help="Start index for behaviors_path data (inclusive)")
    parser.add_argument("--behavior_end_idx", type=int, default=None,
                        help="End index for behaviors_path data (exclusive)")
    parser.add_argument("--behavior_ids_subset", type=str, default='',
                        help="An optional comma-separated list of behavior IDs, or a path to a newline-separated list of behavior IDs. If provided, this will override behavior_start_idx and behavior_end_idx for selecting a subset of behaviors.")
    parser.add_argument("--run_id", type=str, default=None,
                        help="An optional run id for output files, will be appended to output filenames. Must be an integer.")
    parser.add_argument("--overwrite", action="store_true",
                        help="Whether to overwrite existing saved test cases")
    parser.add_argument("--verbose", action="store_true",
                        help="Whether to print out intermediate results")
    
    args = parser.parse_args()
    return args


def main():
    # ========== load arguments and config ========== #
    args = parse_args()
    print(args)
    
    method_name = args.method_name
    experiment_name = args.experiment_name
    save_dir = os.path.abspath(args.save_dir)

    # Load model config file
    config_file = f"configs/model_configs/models.yaml" if not args.models_config_file else args.models_config_file
    with open(config_file) as file:
        model_configs = yaml.full_load(file)

    # Load method config file
    config_file = f"configs/method_configs/{method_name}_config.yaml" if not args.method_config_file else args.method_config_file
    with open(config_file) as file:
        method_configs = yaml.full_load(file)
    
    # Load default method parameters
    if 'default_method_hyperparameters' not in method_configs:
        raise ValueError(f"Default method hyperparameters not found in {config_file}")
    method_config = method_configs.get('default_method_hyperparameters') or {}

    # Update method parameters with experiment parameters
    experiment_config = get_experiment_config(experiment_name, model_configs, method_configs)
    method_config.update(experiment_config)

    if args.run_id:
        assert args.run_id.isdigit(), f"epxected run_id to be an integer; got {args.run_id}"
    
    # ========== generate test cases ========== #
    # load behaviors csv
    with open(args.behaviors_path, 'r', encoding='utf-8') as f:
        reader = csv.DictReader(f)
        behaviors = list(reader)
    if args.behavior_ids_subset:
        print(f"Selecting subset from provided behavior IDs: {args.behavior_ids_subset}")
        if os.path.exists(args.behavior_ids_subset):
            with open(args.behavior_ids_subset, 'r') as f:
                behavior_ids_subset = f.read().splitlines()
        else:
            behavior_ids_subset = args.behavior_ids_subset.split(',')
        behaviors = [b for b in behaviors if b['BehaviorID'] in behavior_ids_subset]
    else:
        behavior_start_idx, behavior_end_idx = args.behavior_start_idx, args.behavior_end_idx
        if args.behavior_start_idx is not None and args.behavior_end_idx is not None:
            print(f"Selecting subset from [{behavior_start_idx}, {behavior_end_idx}]")
            behaviors = behaviors[args.behavior_start_idx:args.behavior_end_idx]

    world_size = torch.cuda.device_count()

    print(f"============= {method_name}, {experiment_name} Config =============")
    print('\n'.join([f"{k}: {v}" for k, v in method_config.items()]))
    print("============= GPU Allocation Configs =============")
    print(f"Total GPUs: {world_size}")
    print("=============")

    # ==== Filtering existed runs =====
    method_class = get_method_class(method_name)

    if not args.overwrite:
        filtered_behaviors = []
        for behavior in behaviors:
            # flatten csv row
            behavior_id = behavior['BehaviorID']
            test_case_file = method_class.get_output_file_path(save_dir, behavior_id, file_type='test_cases', run_id=args.run_id)
            if not os.path.exists(test_case_file):
                filtered_behaviors.append(behavior)
            else:
                print(f"===>Skipping {behavior_id}")
        behaviors = filtered_behaviors
        
        if len(behaviors) == 0:
            print("Found existing test cases for all current behaviors. Exiting.")
            return

    # ==== Init Method ====
    method = init_method(method_class, method_config)

    # ====== Run attack ======
    print("#Behaviors to run:" , len(behaviors))
    test_cases, logs = method.generate_test_cases(behaviors=behaviors, verbose=args.verbose)
    
    # ====== Save test cases and logs ======
    print("Saving outputs to: ", save_dir)
    method.save_test_cases(save_dir, test_cases, logs, method_config=method_config, run_id=args.run_id)
    print("Finished")

if __name__ == "__main__":
    main()