DarkGuardRL
Self-play reinforcement learning for dark pattern detection with Qwen3-4B.
Fine-tuned Qwen3-4B with SFT + GRPO on a custom RL environment for dark pattern detection.
- Designed a Gym-compatible RL simulation environment with discrete, verifiable reward functions for dynamic adversarial decision-making.
- Implemented self-play policy optimization with GRPO, achieving a +30–50 ELO improvement over baseline within 500 training steps.
- Formulated strict multi-objective reward shaping that minimized false positives and eliminated invalid action paths during rollout.
Tech Stack
- Python, PyTorch, Hugging Face TRL, GRPO, Qwen3
Links
- GitHub Repository: Eno_E