online-battle

Online-Battle

Evaluation framework for TurnGate experiments. Use it to run jailbreak attacks, apply defenses, and score results. This codebase is adapted from the original CKA-Agent codebase

Common Commands

python main.py --config config/config_no_defense.yml
python main.py --config config/config_rl_defense.yml
python main.py --config config/config_benchmark_shift_harmbench.yml
export GEMINI_API_KEY="your-key"
python main.py --config config/config_target_shift_gemini.yml
python main.py --config config/config_attacker_shift_maj.yml

Main Areas

• config/ — experiment, method, model, and dataset configs
• methods/ — attack implementations
• defense/ — defense wrappers
• evaluation/ — scoring and metrics
• data/ — datasets used by the experiments

Docs

• config/README.md
• methods/README.md
• scripts/README.md