LFM2.5 Nova Function Calling

Use the LFM2.5-1.2B-Nova-Function-Calling model to convert natural language requests into structured JSON function calls for tools, APIs, and software agents.

Description

This skill leverages a specialized 1.2B parameter Liquid Neural Network fine-tuned for robust function calling. Despite its compact size, it achieves 97% syntax reliability matching GPT-4o class performance, making it ideal for resource-constrained environments while maintaining high-quality structured output generation.

The model excels at:

Instructions

When the user requests function calling or tool invocation capabilities:

1. **Install Dependencies**

- Ensure `transformers`, `torch`, and `unsloth` libraries are installed

- Use `pip install unsloth transformers torch accelerate` if not present

- Verify CUDA availability for GPU acceleration (optional but recommended)

2. **Load the Model**

- Import required modules: `FastLanguageModel` from `unsloth`, `torch`

- Load model using: `NovachronoAI/LFM2.5-1.2B-Nova-Function-Calling-Full`

- Set max sequence length to 4096 tokens

- Enable 4-bit quantization for memory efficiency: `load_in_4bit=True`

- Call `FastLanguageModel.for_inference(model)` to prepare for generation

3. **Format Input Prompts**

- Use ChatML format with `<|im_start|>` and `<|im_end|>` tags

- Structure: `<|im_start|>user\n{query}\n<|im_end|>\n<|im_start|>assistant`

- Include tool definitions in the system message when available

- Keep user queries clear and specific about desired actions

4. **Generate Function Calls**

- Tokenize the prompt with `return_tensors="pt"`

- Move inputs to appropriate device (CUDA/CPU)

- Generate with `max_new_tokens=128` (adjust for complex calls)

- Enable `use_cache=True` for faster inference

- Parse output between `<tool_call>` and `</tool_call>` tags

5. **Parse and Validate Output**

- Extract JSON from `<tool_call>` tags in assistant response

- Validate JSON structure contains `name` and `arguments` fields

- Check that function name matches available tools

- Verify argument types match expected schema

- Handle malformed outputs gracefully with error messages

6. **Execute Tool Calls**

- Map function names to actual tool implementations

- Unpack arguments from JSON into function parameters

- Execute the tool with provided arguments

- Capture and format tool results

- Return results to continue conversation flow

7. **Alternative: GGUF Deployment**

- For local/edge deployment, download GGUF versions from mradermacher repositories

- Standard GGUF: Broad compatibility, general testing

- Imatrix GGUF: Higher quality for low-VRAM devices (recommended)

- Use with llama.cpp, Ollama, or LM Studio for quantized inference

- Select quantization level based on available memory (Q4_K_M recommended baseline)

Performance Characteristics

Example Usage

**Input Query:**

```

I need to calculate the area of a circle with a radius of 5.

```

**Expected Output:**

```json

<tool_call>

{"name": "calculate_circle_area", "arguments": {"radius": 5}}

</tool_call>

```

Constraints

Technical Notes

Model Links