How centerforaisafety/harmbench works
Project type and purpose
• HarmBench is an open-source evaluation framework for automated red teaming of LLMs and robust refusal.
• It provides a standardized, plug-and-play pipeline to generate test cases (attacks), produce model completions, and evaluate the harmfulness of outputs.
File structure
• Top-level:
• scripts/ – orchestration of pipeline steps (run_pipeline.py, shell wrappers)
• baselines/ – all red-teaming method implementations (AutoPrompt, PEZ, TAP, etc.)
• configs/ – YAML definitions for methods, models, and pipeline runs
• data/ – static assets: behavior datasets, classifier validation sets, optimizer targets, images
• multimodalmodels/ – adapters for multimodal LLM APIs (GPT-4V, InstructBLIP, Qwen, etc.)
• adversarial_training/ – optional adversarial fine-tuning code and recipes
• notebooks/ – exploratory analysis and tutorials
• eval_utils.py, generate_completions.py, evaluate_completions.py, generate_test_cases.py, merge_test_cases.py – core Python modules for each pipeline stage
README for architectural clues
• The evaluation pipeline has 4 steps:
1) generate_test_cases
1.5) merge_test_cases
2) generate_completions
3) evaluate_completions
• run_pipeline.py abstracts these steps and can dispatch to Slurm, local, or local_parallel.
• New methods plug in by adding a subfolder in baselines and a matching YAML in configs/method_configs.
• New models plug in by editing configs/model_configs/models.yaml or adding adapter code in multimodalmodels/.
Details to capture
• The plugin interfaces: BaselineMethod (in baseline.py), MultimodalModel class
• The config hierarchy: method_configs → pipeline_configs → model_configs
• Execution modes: slurm vs local vs local_parallel
• Optional adversarial training sub-pipeline under adversarial_training