iromu/Qwen2.5-1.5B-tools
iromu/Qwen2.5-1.5B-tools is a 1.5 billion parameter Qwen2.5-based language model fine-tuned by iromu for tool-calling capabilities. This model leverages LoRA fine-tuning on a specialized dataset to enhance its ability to emit and match tool arguments. It is designed for applications requiring function calling and integration with external tools, offering various GGUF quantizations for optimized performance.
Loading preview...
Model Overview
iromu/Qwen2.5-1.5B-tools is a 1.5 billion parameter model based on the Qwen2.5 architecture, specifically fine-tuned for tool-calling. It is built upon unsloth/Qwen2.5-1.5B-Instruct and utilizes LoRA (dim 32, alpha 32) for efficient adaptation.
Key Capabilities & Training
- Tool-Calling Focus: The model is explicitly fine-tuned to generate tool calls, aiming to improve the accuracy of emitted tool names and arguments.
- Training Data: Fine-tuned on the
sft_toolssplit ofr0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillationwith a sequence length of 4096. - Quantization Options: Available in various GGUF quantizations (e.g., Q4_K_M, Q5_K_M, Q8_0) and NVFP4, providing flexibility for deployment and performance optimization.
Performance Insights
Validation on the sft_tools split shows that while the fine-tuned BF16 model has a lower exact-args match (0.7%) compared to the base (8.8%), certain GGUF quantizations like GGUF-Q5_K_M achieve a 2.2% exact-args match, outperforming other quantized versions and offering improved token generation throughput. For instance, GGUF-Q4_K_M reaches 142.9 tok/s, significantly faster than the BF16 version's 38.0 tok/s.
Use Cases
This model is particularly suited for applications where function calling and interaction with external APIs or tools are required. Its optimized tool-calling fine-tune makes it a candidate for building agents or systems that need to interpret user requests and translate them into structured tool invocations.