dinushiTJ/action-gemma-2-2b-it-gguf

ActionGemma 2B: GGUF (run it locally)

GGUF builds of ActionGemma 2B, a Gemma 2 2B model fine-tuned for function calling. Use these files to run it on a laptop or CPU with llama.cpp, Ollama or LM Studio. No cloud API is needed, and your data stays on your machine.

Compared with the base model, ActionGemma reaches 93.74% macro F1 (+7.82%), cuts hallucinated function calls from 16% to 6.7%, and uses about 17% fewer tokens. See the main model card for full results, training details and limitations.

Files

Run it

# Ollama
ollama run hf.co/dinushiTJ/action-gemma-2-2b-it-gguf:Q4_K_M

# llama.cpp
llama-cli -hf dinushiTJ/action-gemma-2-2b-it-gguf:Q4_K_M

Prompt format

The model was trained with the prompt below. Put it at the start of the first user message, with your function definitions (JSON Schema) inside the <functions> tags:

You have access to the following functions enclosed within <functions></functions> tags. Use them if required.
<functions>
{"name": "get_weather", "description": "Get the current weather in a given location",
 "parameters": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}}
</functions>

To call a function, respond using the following format:
<functioncall> {"name": "function_name", "arguments": {"arg_1": "value_1", "arg_2": "value_2", ...}} </functioncall>

If no available function matches the user request, respond politely that you cannot assist with the request.

The model replies with <functioncall> {...} </functioncall>, which is easy to parse with a regex.

Related

• —LoRA adapter and full model card: `dinushiTJ/action-gemma-2-2b-it-lora`
• —Merged FP16 weights for vLLM: `dinushiTJ/action-gemma-2-2b-it-vllm-f16`
• —Collection with the training and evaluation datasets

Use of this model is subject to the Gemma Terms of Use. Author: Dinushi Jayasinghe.