brokencircuitranch/gemma4-hermes-tools
The brokencircuitranch/gemma4-hermes-tools is a 26 billion parameter Mixture of Experts (MoE) model, fine-tuned from google/gemma-4-26B-A4B-it. It specializes in reliable tool use and function calling, making it suitable for agentic pipelines. This model excels at generating structured tool calls based on instruction following and reasoning tasks.
Loading preview...
brokencircuitranch/gemma4-hermes-tools Overview
This model is a 26 billion parameter Mixture of Experts (MoE) variant, fine-tuned from the google/gemma-4-26B-A4B-it base model. Its primary focus is on enhancing reliable tool use and function calling capabilities, making it particularly effective for integration into agentic workflows.
Key Capabilities
- Structured Tool Call Generation: Optimized to produce consistent and accurate structured tool calls, crucial for automated agents.
- Instruction Following: Benefits from additional training on general instruction following and reasoning examples.
- Efficient Fine-tuning: Utilized QLoRA (4-bit) on NVIDIA A100 hardware, then merged to 16-bit weights for deployment.
Training Details
The model was fine-tuned using the Unsloth framework on a combined dataset of nearly 7,000 examples. This included 1,893 examples from NousResearch/hermes-function-calling-v1 for structured tool use, and 5,000 sampled examples from teknium/OpenHermes-2.5 for broader instruction following. The training achieved a final loss of 0.224.
Intended Use
This model is specifically designed for agentic pipelines that require robust and predictable generation of structured tool calls. It has been tested for local inference using Ollama, indicating its suitability for various deployment environments.