Marco-Danz/veneto-mistral-dataset

Veneto Mistral Dataset

A conversational dataset for training AI models in Venetian language (vèneto).

Description

This dataset was created to preserve and digitalize the Venetian language through artificial intelligence. It contains approximately 11,800 examples of conversations, texts, and translations in Venetian language, extracted from authentic sources and validated for linguistic quality.

The dataset was specifically designed for fine-tuning large language models (LLMs) like Mistral 7B, with the goal of creating conversational assistants capable of:

• —Conversing naturally in Venetian
• —Translating between Italian and Venetian
• —Providing information about Venetian culture
• —Generating authentic text in Venetian style

This is an experimental cultural preservation project that represents the first step towards creating AI tools for regional Italian languages that are underrepresented in the technological landscape.

Features

• —Dataset Size: 10,655 training examples + 1,184 validation examples (11,839 total)
• —Format: Chat format with user/assistant messages in conversational style
• —Quality: Texts extracted from authentic sources (Venetian Wikipedia, Goldoni's theatre) and manually validated for conversational examples
• —Total Size: ~5MB of pure Venetian text
• —Split Ratio: 90% training / 10% validation

Composition

Sources

The dataset is composed of three main sources:

1. 1.Wikipedia in Venetian (vec.wikipedia.org) - ~11,000 articles
2. 2.Encyclopedic articles on history, geography, culture, and sciences
3. 3.Text written in standard Venetian (predominantly Venetian dialect)
4. 4.Largest source of the dataset (~93% of content)

1. 1.Carlo Goldoni's Theatre - "Il servitore di due padroni" (The Servant of Two Masters)
2. 2.Authentic theatrical dialogues from the 18th century
3. 3.Examples of natural conversation in historical context
4. 4.Important for authentic conversational style

1. 1.Manually created conversational examples
2. 2.Modern conversations created by native speakers
3. 3.Cover everyday and contemporary situations
4. 4.Validated for naturalness and authenticity

Note on regional representation: The dataset contains mainly standard Venetian (Venetian dialect). Specific regional variants (Paduan, Veronese, Trevisan, etc.) may be underrepresented.

Split

• —Training: 10,655 examples (90%)
• —Validation: 1,184 examples (10%)

The split was performed in a stratified manner to maintain the distribution of sources in both sets.

Format

Each example in the dataset follows the conversational chat format with structured messages:

{
  "messages": [
    {
      "role": "user",
      "content": "Cossa seto de Africa Orientałe?"
    },
    {
      "role": "assistant",
      "content": "Ł'Africa orientałe ła xe ła porsion del continente african che ła se estende giograficamente pì a est."
    }
  ]
}

Examples include:

• —Questions and answers on encyclopedic topics
• —Daily conversations
• —Translation requests Italian ↔ Venetian
• —Cultural explanations

Usage

Installation

pip install datasets

Loading Dataset

from datasets import load_dataset

# Load dataset
dataset = load_dataset("Marco-Danz/veneto-mistral-dataset")

# Access train and validation
train_data = dataset['train']
val_data = dataset['validation']

# Print first example
print(train_data[0])

Preprocessing for Training

The dataset is already formatted for use with Mistral and similar models that use the [INST]...[/INST] format:

def format_for_mistral(example):
    """Format an example for Mistral."""
    messages = example['messages']

    # Mistral format: [INST] user_message [/INST] assistant_response
    user_msg = messages[0]['content']
    assistant_msg = messages[1]['content']

    return f"[INST] {user_msg} [/INST] {assistant_msg}"

# Apply formatting
formatted = train_data.map(lambda x: {"text": format_for_mistral(x)})

Example Usage in Training

from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset

# Load dataset and tokenizer
dataset = load_dataset("Marco-Danz/veneto-mistral-dataset")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.3")

# Tokenize dataset
def tokenize(example):
    text = format_for_mistral(example)
    return tokenizer(text, truncation=True, max_length=2048)

tokenized_dataset = dataset.map(tokenize, remove_columns=dataset['train'].column_names)

# Use for training with Trainer or custom loop

Statistics

Dataset Composition by Source

• —Wikipedia Veneta: ~93% (~11,000 examples)
• —Goldoni's Theatre: ~5% (~600 examples)
• —Conversational examples: ~2% (~200 examples)

Average Lengths

• —Average user message: ~50 tokens
• —Average assistant message: ~150 tokens
• —Max sequence length: 2048 tokens (truncated)

Topics Covered

• —Venetian geography and history
• —Culture and traditions
• —Art and literature
• —Science and technology
• —Daily conversations
• —Italian-Venetian translation

License

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

You are free to:

• —Share: Copy and redistribute the material in any medium or format
• —Adapt: Remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• —Attribution: You must give appropriate credit and indicate if changes were made
• —ShareAlike: If you remix, transform, or build upon the material, you must distribute your contributions under the same license

Source Licenses

• —Wikipedia content: CC BY-SA 3.0 (compatible with CC BY-SA 4.0)
• —Goldoni's works: Public domain (18th century)
• —Manual examples: CC BY-SA 4.0

Limitations and Ethical Considerations

Limitations

• —Limited size: ~11,800 examples is a relatively small dataset compared to mainstream datasets (which have millions of examples)
• —Regional variance: Mainly standard Venetian (Venetian dialect), other regional variants may be underrepresented
• —Domain coverage: Limited mainly to encyclopedic and basic conversational content
• —Historical language: Includes 18th-century text (Goldoni) which may differ from modern Venetian
• —Written vs Spoken: Predominantly written text, some aspects of colloquial speech may be missing

Ethical Considerations

• —Cultural preservation: This dataset is created with the intent of preserving a minority regional language
• —Representation: Does not claim to represent all variants of Venetian, which vary significantly between provinces
• —Quality validation: Conversational examples have been validated mostly automatically they may contain errors or inaccuracies
• —Intended use: Designed for educational, cultural, and research use - not for certified professional translations

Biases

• —Geographical bias towards Venetian dialect
• —Potential historical bias from 18th-century theatre
• —Limited modern colloquial expressions

Citation

If you use this dataset in your research, please cite:

@misc{veneto-mistral-dataset-2025,
  author = {Marco Danzo},
  title = {Veneto Mistral Dataset: A Conversational Dataset for Venetian Language},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/Marco-Danz/veneto-mistral-dataset}}
}

Related Resources

Models Trained on This Dataset

• —Marco-Danz/leoncino-mistral-7b-veneto - Mistral 7B fine-tuned with LoRA

Related Projects

• —Repository: veneto-mistral-finetuning - Complete code for training and inference
• —Venetian Wikipedia: vec.wikipedia.org

Contact

• —Email: danzo.marco1@gmail.com
• —GitHub: https://github.com/marco25d
• —LinkedIn: Marco Danzo

Contributions

🤝 Contributions welcome!

This is an open cultural preservation project. Particularly appreciated contributions:

• —📚 Dataset expansion: Add more texts in Venetian from authentic sources
• —🗣️ Regional variants: Contributions to better represent Paduan, Veronese, Trevisan, etc.
• —✍️ Modern conversational examples: Contemporary dialogues validated by native speakers
• —🔍 Quality review: Report and correct linguistic errors
• —📖 Documentation: Improvements to documentation and usage examples

To contribute:

1. 1.Fork the GitHub repository
2. 2.Add your data or corrections
3. 3.Open a Pull Request with detailed description
4. 4.Or open an Issue for discussions and proposals

Acknowledgments

This dataset was made possible thanks to:

• —Venetian Wikipedia and all contributors who created encyclopedic content in Venetian
• —Carlo Goldoni for his immortal theatrical texts that preserve 18th-century Venetian
• —Venetian community for actively preserving the language
• —Mistral AI for making the base model available open source
• —HuggingFace for the infrastructure for sharing datasets and models

Special thanks to everyone working to preserve Italian regional languages.

Note: This dataset is part of a larger project for preservation and digitalization of the Venetian language through artificial intelligence. It represents a starting point for future more complete and sophisticated developments.

Thank you to anyone who uses this dataset and contributes to the preservation of the Venetian language! 🙏

Viva el vèneto! 🎭🦁