Model Overview

Seungyoun/qwen2.5-3b-it_searchR1-like-multiturn is a 3.1 billion parameter instruction-tuned model built upon the Qwen 2.5-3B-instruct architecture. Developed by Seungyoun Shin, this model is a re-implementation of the Search-R1 approach, specifically designed for multi-turn tool-calling capabilities. It was trained using GRPO via the open-source VERL framework on the nq-hotpotqa-train dataset.

Key Capabilities

• Integrated Web Search: The model can autonomously call a search function (e.g., DuckDuckGo) to retrieve external information, parsing tool calls and integrating search results into its reasoning process.
• Multi-turn Reasoning: It is designed to handle complex, multi-turn interactions, performing chain-of-thought reasoning within <think> blocks before generating answers or making tool calls.
• Enhanced Question Answering: The model demonstrates improved performance on several question-answering datasets, including NQ, TriviaQA, PopQA, HotpotQA, 2Wiki, and Bamboogle, often surpassing the original Search-R1 implementation.
• Structured Output: It follows a specific reasoning style, emitting JSON-formatted tool calls and providing concise answers within <answer> blocks.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

• Complex Question Answering: Answering questions that necessitate external knowledge retrieval and multi-step reasoning.
• Tool-Augmented LLM Applications: Scenarios where an LLM needs to interact with external tools or APIs to gather information.
• Research and Development: As a strong baseline or starting point for further research into tool-calling, reasoning, and knowledge-augmented language models, especially within the Qwen 2.5 ecosystem.