Tucan-BG-Eval-v1.0 🇧🇬

Paper: https://arxiv.org/abs/2506.23394

Official evaluation dataset for Tucan models - a series of Bulgarian language models fine-tuned for function calling and tool use.

📊 Dataset Overview

This dataset contains 120 curated samples designed to evaluate function-calling capabilities in Bulgarian language models. Each sample includes:

• User queries in Bulgarian
• Available functions with their parameters
• Expected behavior (function calls or text responses)
• Evaluation criteria
• Scenario types for different use cases

🎯 Evaluation Scenarios

The dataset covers various function-calling scenarios:

• function_call_required - Cases where a function must be called
• multiple_functions_selection - Choosing between multiple available functions
• irrelevant_question_with_functions - Questions that should be answered without function calls
• no_functions_available - General knowledge questions
• missing_required_parameters - Cases requiring parameter clarification

🔧 Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("s-emanuilov/Tucan-BG-Eval-v1.0")

# Access the samples
samples = dataset["train"]

For detailed evaluation instructions, see the Tucan Evaluation Framework.

📝 Dataset Format

Each sample contains:

{
    "user_message": "Bulgarian user query",
    "functions": [
        {
            "name": "function_name",
            "description": "Function description",
            "parameters": {
                "properties": {...},
                "required": [...]
            }
        }
    ],
    "expected_behavior": {
        "expected_function": "function_name",
        "expected_parameters": {...},
        "should_call_function": true/false
    },
    "evaluation_criteria": {...},
    "scenario_type": "scenario_category"
}

🔗 Links

• Tucan Models Collection
• Evaluation Framework