wckwan/Search-R1-Qwen3-8B-SDAR
Search-R1-Qwen3-8B-SDAR is an 8 billion parameter Qwen3-based language model developed by wckwan, fine-tuned with GRPO on Search-R1 multi-turn retrieval QA. This model incorporates SDAR self-distillation to enhance its performance in tool-call interactions. It is specifically designed for multi-turn retrieval question answering tasks, leveraging its specialized training for improved conversational search capabilities.
Loading preview...
Model Overview
wckwan/Search-R1-Qwen3-8B-SDAR is an 8 billion parameter model built upon the Qwen3-8B architecture. It has been specifically fine-tuned using GRPO (Generalized Reinforcement Learning with Policy Optimization) on the Search-R1 multi-turn retrieval QA dataset. A key differentiator is its use of SDAR (Self-Distillation with Adversarial Regularization), which contributes to its specialized performance.
Key Capabilities
- Multi-turn Retrieval QA: Optimized for complex question-answering scenarios involving multiple conversational turns and retrieval of information.
- Tool-Call Interaction Format: Trained to understand and generate responses within a tool-call interaction framework, making it suitable for applications requiring external tool usage.
- Policy Training: Developed as a policy model, indicating its suitability for decision-making and sequence generation in interactive systems.
Training Details
The model underwent 400 training steps, focusing on the Search-R1 tool-call interaction format. Intermediate checkpoints are available, allowing for analysis of training progression or resuming training from specific stages. Raw FSDP checkpoints are also provided for advanced users wishing to resume training.
Good For
- Developing conversational AI agents that require multi-turn information retrieval.
- Applications needing models capable of handling tool-call interactions.
- Research into reinforcement learning and self-distillation techniques for QA models.