beyoru/Qwen3-4B-I-1209
The beyoru/Qwen3-4B-I-1209 model is a 4-billion parameter instruction-tuned language model, based on Qwen3-4B-Instruct-2507, developed by Beyoru. It is specifically fine-tuned using Reinforcement Learning (GRPO) with multiple reward functions to excel in tool-use and function call generation. This model demonstrates enhanced performance in tool-use scenarios, achieving an overall accuracy of 0.7233 on ACEBench, outperforming its base model and other similar-sized models.
Loading preview...
Model Overview
The beyoru/Qwen3-4B-I-1209 is a 4-billion parameter model, building upon the Qwen3-4B-Instruct-2507 base. Developed by Beyoru, this model is uniquely fine-tuned for tool-use and function call generation using Reinforcement Learning (GRPO).
Key Differentiators
- Specialized Fine-tuning: Utilizes GRPO with a multi-signal reward system, including rule-based, self-certainty, and tool-call rewards, to optimize for accurate and confident function calling.
- Enhanced Tool-Use Performance: Achieves an overall accuracy of 0.7233 on the ACEBench evaluation, significantly surpassing its base model (
Qwen3-4B-Instruct-2507at 0.635) andSalesforce/Llama-xLAM-2-8b-fc-r(0.5792).
Training Configuration
The model was trained with an AdamW optimizer, a learning rate of 5e-6 with cosine decay, and a cosine_with_min_lr scheduler, generating 4 responses per prompt.
Ideal Use Cases
This model is particularly well-suited for applications requiring:
- Reliable Function Calling: Generating correct function names and arguments for external tools.
- Automated Workflow Integration: Systems that need to interact with APIs or other software components through function calls.
- Agentic AI Systems: Powering agents that can autonomously decide and execute actions via tool invocation.