TUCAN (Tool-Using Capable Assistant Navigator) is a series of open-source Bulgarian language models fine-tuned specifically for function calling and tool use.

These models can interact with external tools, APIs, and databases, making them appropriate for building AI agents and Model Context Protocol (MCP) applications.

Built on top of BgGPT models from INSAIT Institute, which were themselves built on Gemma 2, Tucan models have been enhanced with function-calling capabilities.

Motivation 🎯

Although BgGPT models demonstrate strong Bulgarian language comprehension, they face challenges in maintaining the precise formatting necessary for consistent function calling. Despite implementing detailed system prompts, their performance in this specific task remains suboptimal.

This project addresses that gap by fine-tuning BgGPT, providing the Bulgarian AI community with proper tool-use capabilities in their native language.

Models and variants 📦

Available in three sizes with full models, LoRA adapters, and quantized GGUF variants:

Model Size	Full Model	LoRA Adapter	GGUF (Quantized)
2.6B	Tucan-2.6B-v1.0	LoRA	GGUF
9B	Tucan-9B-v1.0 📍	LoRA	GGUF
27B	Tucan-27B-v1.0	LoRA	GGUF

GGUF variants include: q4_k_m, q5_k_m, q6_k, q8_0, q4_0 quantizations

📍 Current model/repo

Models and quantizations are also available for easy use in Ollama: https://ollama.com/s_emanuilov/tucan

Benchmarks 📊

All evaluations were performed using the Tucan evaluation framework, with results averaged across multiple runs. Tucan models demonstrate superior function-calling capabilities compared to their BgGPT counterparts, with particularly strong improvements in smaller model sizes. To ensure no catastrophic forgetting occurred, we evaluated knowledge retention using EleutherAI's lm-evaluation-harness on Bulgarian benchmarks, confirming that each Tucan model maintains performance on par with its BgGPT equivalent.

Model	Function Calling	HellaswagBG	WinograndeBG	ARC-Easy-BG	ARC-Challenge-BG
Tucan-2.6B-v1.0 🔥	0.7875	0.5924	0.6456	0.5657	0.3754
Tucan-9B-v1.0 🔥	0.8667	0.7046	0.7151	0.7024	0.5188
Tucan-27B-v1.0 🔥	0.875	0.6179	0.6275	0.6486	0.442
BgGPT-Gemma-2-2.6B-IT-v1.0	0.5874	0.6306	0.5821	0.5657	0.372
BgGPT-Gemma-2-9B-IT-v1.0	0.7833	0.7057	0.719	0.7231	0.5188
BgGPT-Gemma-2-27B-IT-v1.0	0.8667	0.62	0.6212	0.6587	0.459

Note: 27B models were evaluated in 8-bit precision for comparison purposes.

Usage 🛠️

Quick start ⚡

pip install -U "transformers[torch]" accelerate bitsandbytes

Prompt format ⚙️

Critical: Use this format for function calling for the best results.

📋 Required system prompt template

<bos><start_of_turn>user
Ти си полезен AI асистент, който предоставя полезни и точни отговори.

Имаш достъп и можеш да извикаш една или повече функции, за да помогнеш с потребителското запитване. Използвай ги, само ако е необходимо и подходящо.

Когато използваш функция, форматирай извикването ѝ в блок ```tool_call``` на отделен ред, a след това ще получиш резултат от изпълнението в блок ```toll_response```.

## Шаблон за извикване: 
```tool_call
{"name": <function-name>, "arguments": <args-json-object>}```

## Налични функции:
[your function definitions here]

## Потребителска заявка: 
[your query in Bulgarian]<end_of_turn>
<start_of_turn>model

Note 📝

The model only generates the tool_call blocks with function names and parameters - it doesn't actually execute the functions. Your client application must parse these generated calls, execute the actual functions (API calls, database queries, etc.), and provide the results back to the model in tool_response blocks for the conversation to continue the interperation of the results. A full demo is comming soon.

Python example 🐍

💻 Complete Working Example

python

import torch
import json
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig

# Load model
model_name = "s-emanuilov/Tucan-2.6B-v1.0"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="eager"  # Required for Gemma models
)

# Create prompt with system template
def create_prompt(functions, user_query):
    system_prompt = """Ти си полезен AI асистент, който предоставя полезни и точни отговори.

Имаш достъп и можеш да извикаш една или повече функции, за да помогнеш с потребителското запитване. Използвай ги, само ако е необходимо и подходящо.

Когато използваш функция, форматирай извикването ѝ в блок ```tool_call``` на отделен ред, a след това ще получиш резултат от изпълнението в блок ```toll_response```.

## Шаблон за извикване: 
```tool_call
{{"name": <function-name>, "arguments": <args-json-object>}}```
"""
    
    functions_text = json.dumps(functions, ensure_ascii=False, indent=2)
    full_prompt = f"{system_prompt}\n## Налични функции:\n{functions_text}\n\n## Потребителска заявка:\n{user_query}"
    
    chat = [{"role": "user", "content": full_prompt}]
    return tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

# Example usage
functions = [{
    "name": "create_calendar_event",
    "description": "Creates a new event in Google Calendar.",
    "parameters": {
        "type": "object",
        "properties": {
            "title": {"type": "string"},
            "date": {"type": "string"},
            "start_time": {"type": "string"},
            "end_time": {"type": "string"}
        },
        "required": ["title", "date", "start_time", "end_time"]
    }
}]

query = "Създай събитие 'Годишен преглед' за 8-ми юни 2025 от 14:00 до 14:30."

# Generate response
prompt = create_prompt(functions, query)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.1,
    top_k=25,
    top_p=1.0,
    repetition_penalty=1.1,
    do_sample=True,
    eos_token_id=[tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<end_of_turn>")],
    pad_token_id=tokenizer.eos_token_id
)

result = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(result)

Performance & Dataset 📊

📄 Full methodology, dataset details, and comprehensive evaluation results coming in the upcoming paper

Dataset: 10,000+ bilingual (Bulgarian/English) function-calling examples across 1,000+ topics, including tool calls with single/multiple arguments, optional parameters, follow-up queries, multi-tool selection, ambiguous queries requiring clarification, and conversational interactions without tool use. Data sourced from manual curation and synthetic generation (Gemini Pro 2.5/GPT-4.1/Sonnet 4).

Results: Significant improvements in tool-use capabilities over base BgGPT models: 34.1% for 2.6B, 10.6% for 9B, and 1.0% for 27B models in internal benchmarks. Beyond raw function-calling scores, all Tucan models demonstrate more natural conversational flow while maintaining tool-use capabilities, retaining their base knowledge.

Acknowledgments 🙏

Built on top of BgGPT series.

Questions & Contact 💬

For questions, collaboration, or feedback: Connect on LinkedIn

llm-bg/Tucan-9B-v1.0

Tucan-9B-v1.0

Bulgarian Language Models for Function Calling 🇧🇬

Overview 🚀