functiongemma-agent-gguf

FunctionGemma Agent GGUF

A fine-tuned version of FunctionGemma-270M for agentic tool-calling tasks, converted to GGUF format for use with llama.cpp and llama-agent.

Model Details

Training

Fine-tuned using Unsloth with LoRA on HuggingFace Jobs infrastructure.

Training Configuration:

• LoRA rank: 128, alpha: 256
• Epochs: 3
• Learning rate: 2e-4
• Batch size: 4, gradient accumulation: 2
• Hardware: NVIDIA A100-80GB
• Training method: SFT with train_on_responses_only

Dataset: 7,500 synthetic examples covering:

• Multi-step tool chaining (glob → read → edit)
• Error recovery patterns
• Clarification dialogs
• No-tool responses
• Parallel tool calls

Tools

The model is trained on 5 tools matching llama-agent:

Usage

With llama.cpp

# Download
wget https://huggingface.co/victor/functiongemma-agent-gguf/resolve/main/functiongemma-270m-it.Q4_K_M.gguf

# Run inference
./llama-cli -m functiongemma-270m-it.Q4_K_M.gguf -p "<start_of_turn>user
Read the main.py file
<end_of_turn>
<start_of_turn>model"

With llama-agent

./llama-agent -m functiongemma-270m-it.Q4_K_M.gguf

Format

Uses FunctionGemma's native format with <escape> delimiters:

<start_of_turn>user
Fix the typo in config.json
<end_of_turn>
<start_of_turn>model
<think>I need to find and read the config file first.</think>
<start_function_call>call:glob{pattern:<escape>**/config.json<escape>}<end_function_call>
<end_of_turn>
<start_of_turn>developer
<start_function_response>response:glob{stdout:<escape>src/config.json<escape>,stderr:<escape><escape>,exit_code:0}<end_function_response>
<end_of_turn>
...

License

This model inherits the Gemma license from the base model.

Links

• Training script: victor/llama-agent-training
• Dataset: victor/functiongemma-agent-sft
• llama-agent: github.com/ggml-org/llama.cpp/tools/agent