Hugging Face Overview
FunReason-MT-4B: Advanced Multi-Turn Function Calling
FunReason-MT-4B is a 4 billion parameter Large Language Model (LLM) developed by Bingguang, specifically fine-tuned for advanced multi-turn Function Calling (FC) and agentic tool-use. Built upon the Qwen3-4B-Instruct-2507 base model, it leverages a novel FunReason-MT data synthesis framework to generate high-quality, complex multi-turn data.
Key Capabilities & Performance
- Superior Multi-Turn Function Calling: Achieves 57.75% on the BFCLv3 Multi-Turn benchmark, significantly outperforming its base model (15.75%) and competitive with Claude-Sonnet-4.
- Leading Agentic Tool-Use: Ranks highest with 15.10% on the BFCLv4 Out-of-Distribution (OOD) Agentic Evaluation, excelling in tasks like Web Search and Memory.
- Data Synthesis Innovation: Utilizes a three-phase framework for data generation, focusing on environment-API graph interactions, advanced tool-query synthesis, and a guided iterative Chain-of-Thought (CoT) process.
Training Details
The model was fine-tuned using Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) on 16,000 high-quality multi-turn samples. This training was conducted on 32 NVIDIA H20 GPUs, utilizing LLama-Factory and Verl libraries.
Ideal Use Cases
- Developing AI agents requiring complex, multi-step interactions with tools.
- Applications demanding robust function calling capabilities in conversational or multi-turn scenarios.
- Research into advanced data synthesis techniques for improving LLM performance on intricate tasks.