LLMBG ToolUse 9B Bulgarian Function Calling

A 9B parameter Bulgarian language model optimized for function calling, tool use, and Model Context Protocol (MCP) integration. This is a quantized GGUF version of the Tucan-9B-v1.0 model with weighted imatrix quantization for optimal performance across different hardware configurations.

What This Model Does

This model specializes in:

Instructions for AI Agents

When using this model for tool-enabled tasks:

1. **Model Selection**

- Choose an appropriate quantization level based on available hardware

- Recommended: Q4_K_M (5.9GB) for balanced speed and quality

- For resource-constrained environments: Q4_K_S (5.6GB) for optimal size/speed/quality

- For maximum quality: Q6_K (7.7GB) for near-original model performance

2. **Loading the Model**

- Download the desired GGUF quantization from HuggingFace

- Load using llama.cpp, Ollama, LM Studio, or other GGUF-compatible inference engines

- Configure context window and token limits based on your use case

3. **Function Calling Setup**

- Define your tools/functions with clear JSON schemas

- Use structured prompts that specify available tools

- Format function definitions with parameter types and descriptions

- Include examples of expected function call format

4. **Prompting for Tool Use**

- Be explicit about available tools at the start of conversations

- Use Bulgarian or multilingual prompts as appropriate

- Request function calls in structured format (JSON preferred)

- Provide clear context about when tools should be invoked

5. **MCP Integration**

- Configure MCP protocol endpoints if using agent frameworks

- Map model function calls to MCP tool invocations

- Handle tool execution results and feed back to model

- Implement error handling for failed tool executions

6. **Best Practices**

- Start with higher quantization levels and reduce if needed for performance

- Monitor token usage and context window utilization

- Validate function call outputs before execution

- Provide clear feedback loops between tool results and model responses

- Test Bulgarian language capabilities with native prompts for optimal performance

Quantization Levels Reference

| Quantization | Size | Use Case |

|--------------|------|----------|

| IQ3_XXS - Q3_K | 3.9-5.2GB | Lower quality, very constrained resources |

| Q4_K_S | 5.6GB | Optimal size/speed/quality balance |

| Q4_K_M | 5.9GB | Fast, recommended for most use cases |

| Q5_K_M | 6.7GB | Higher quality, moderate resource usage |

| Q6_K | 7.7GB | Near-original quality, higher resource usage |

Example Usage

```python

Load model with llama-cpp-python

from llama_cpp import Llama

llm = Llama(

model_path="LLMBG-ToolUse-9B-v1.0.i1-Q4_K_M.gguf",

n_ctx=4096,

n_threads=8

)

Define tools

tools = [{

"name": "get_weather",

"description": "Get current weather for a location",

"parameters": {

"type": "object",

"properties": {

"location": {"type": "string"}

}

}

}]

Prompt with tool context

prompt = f"""Available tools: {tools}

User: Какво е времето в София? (What's the weather in Sofia?)