phanviethoang1512/Qwen3-8B-CSD-Agent
phanviethoang1512/Qwen3-8B-CSD-Agent is an 8 billion parameter Qwen3-based agent model, fine-tuned using multi-turn GRPO and Coached On-Policy Self-Distillation (CSD) with MIMESIS-9B as the user simulator. This model excels in interactive agent environments by leveraging token-level supervision from a coach, achieving a mean task score of 31.09 across eight UserRL environments. It is specifically designed for developing and evaluating interactive agents in complex dialogue and task-oriented scenarios.
Loading preview...
Overview
phanviethoang1512/Qwen3-8B-CSD-Agent is an 8 billion parameter model based on the Qwen3 architecture, specifically developed as a CSD (Coached On-Policy Self-Distillation) agent. It was trained using multi-turn GRPO combined with Coached On-Policy Self-Distillation, utilizing MIMESIS-9B as the training user simulator. This approach allows a "coach" to provide concise notes on the agent's responses, generating dense, token-level supervision that goes beyond sparse task rewards.
Key Capabilities
- Enhanced Interactive Agent Performance: Achieves a mean task score of 31.09 across eight UserRL environments, outperforming other Qwen3-8B agent variants like Base (18.49), UserRL (26.10), InfoPO (26.65), and UserRL+ (29.54).
- Advanced Training Methodology: Leverages Coached On-Policy Self-Distillation (CSD) for more effective learning in interactive settings, providing token-level feedback.
- Robust Evaluation: Demonstrated superior performance across nine evaluation user models not used in training, including held-out environments like IntentionGym, TelepathyGym, and SearchGym.
Good for
- Developing and evaluating interactive agents in complex, multi-turn dialogue scenarios.
- Researching advanced reinforcement learning techniques for language agents.
- Applications requiring agents that can learn from detailed, token-level feedback rather than just sparse rewards.