LLMBG ToolUse 27B Bulgarian Model

A 27B parameter Bulgarian language model fine-tuned for function calling, Model Context Protocol (MCP), and tool use capabilities. This is a quantized GGUF version of llm-bg/Tucan-27B-v1.0 with weighted imatrix quantization for optimal size/quality tradeoff.

What This Skill Does

This skill provides access to a Bulgarian-language AI model that excels at function calling and tool use. It's particularly useful for building AI agents and assistants that need to work with Bulgarian language content while maintaining strong tool-calling capabilities.

Instructions for AI Agent

When a user requests to use this Bulgarian tool-use model:

1. **Verify Model Availability**

- Check if the user has llama.cpp, Ollama, or another GGUF-compatible inference engine installed

- Confirm the model file is downloaded or provide instructions to download it

2. **Select Appropriate Quantization**

- For systems with limited RAM (8-16GB): Recommend IQ3_S or Q4_K_S quantization (~12-16GB)

- For systems with moderate RAM (16-32GB): Recommend Q4_K_M or Q5_K_M (~17-20GB)

- For systems with ample RAM (32GB+): Recommend Q6_K for best quality (~22GB)

- For extremely constrained systems: IQ2_S or IQ2_M can work but with quality tradeoffs

3. **Download Instructions**

- Provide the appropriate HuggingFace download URL for the selected quantization

- Example: `https://huggingface.co/mradermacher/LLMBG-ToolUse-27B-v1.0-i1-GGUF/resolve/main/LLMBG-ToolUse-27B-v1.0.i1-Q4_K_M.gguf`

4. **Configuration for Tool Use**

- Set temperature to 0.1-0.3 for function calling (requires precision)

- Enable tool/function calling mode in the inference engine

- Configure MCP settings if using Model Context Protocol

- Set context window appropriately (model supports extended context)

5. **Usage Examples**

- Demonstrate how to define functions/tools for the model

- Show Bulgarian language prompts that trigger tool use

- Provide examples of MCP integration if applicable

6. **Performance Optimization**

- Enable GPU offloading if available (Metal, CUDA, ROCm)

- Adjust batch size based on available VRAM

- Monitor response quality and adjust quantization if needed

7. **Bulgarian Language Considerations**

- This model is optimized for Bulgarian language input/output

- Function definitions can be in English, but prompts should be in Bulgarian for best results

- The model understands multilingual contexts but performs best with Bulgarian

Recommended Quantizations

Important Notes

Example Usage

```bash

Download with wget

wget https://huggingface.co/mradermacher/LLMBG-ToolUse-27B-v1.0-i1-GGUF/resolve/main/LLMBG-ToolUse-27B-v1.0.i1-Q4_K_M.gguf

Run with llama.cpp

./llama-cli -m LLMBG-ToolUse-27B-v1.0.i1-Q4_K_M.gguf -p "Какви функции мога да използвам?" --temp 0.2

Run with Ollama (after creating Modelfile)

ollama run llmbg-tooluse

```

Constraints