uvia1.3

UVIA v1.3 - Specialized Language Model for Brazilian Viticulture and Enology

UVIA is a specialized language model for Brazilian viticulture, enology, and wine production. Developed through incremental fine-tuning and strategic rebalancing, UVIA v1.3 achieves perfect scores (1.0) in Brazilian focus and professional structure metrics.

🌟 Key Features

• 🏆 Brazilian Specialization: Perfect score (1.0) in Brazilian viticulture expertise
• ⚡ Edge Computing: Designed for offline operation in remote vineyards
• 💰 Zero Operational Cost: Free inference after initial setup
• 🔒 Total Privacy: Local data processing, no cloud dependency
• 🧠 Agriculture 4.0 Ready: Native IoT integration and agentified systems
• 📈 Superior Performance: 138% improvement over previous versions

📊 Benchmark Results

Local Benchmarks (vs Qwen3-8B)

Rebalancing Results (v1.1 → v1.3)

• Overall Score: 0.420 → 1.000 (+138%)
• Brazilian Focus: 0.60 → 1.00 (+67%)
• Structure: 0.00 → 1.00 (+∞%)

OpenAI Models Comparison

🚀 Use Cases

Brazilian Viticulture Applications

• Technical Consulting: Specialized advice for Brazilian winemakers
• Professional Education: Training materials for enologists and viticulturists
• Market Analysis: Brazilian wine industry insights
• Regulatory Guidance: IN 5/2010, IN 12/2010 compliance assistance

Agriculture 4.0 Features

• Edge Computing: Offline operation in remote vineyards
• IoT Integration: Direct connection with agricultural sensors
• Autonomous Systems: Agentified decision-making for vineyards
• Remote Monitoring: Vineyard analysis without internet connectivity

💻 Usage

With Ollama (Recommended)

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Run UVIA v1.3
ollama run uvia-1-3

# Example query
"What are the best grape varieties for Brazilian tropical climates?"

With Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model and tokenizer
model_name = "uvia-br/UVIA-v1.3"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Generate response
input_text = "Como identificar problemas de fermentação em vinhos brasileiros?"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=512, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

With Ollama API

import requests

response = requests.post('http://localhost:11434/api/generate',
    json={
        "model": "uvia-1-3",
        "prompt": "What are the main challenges in Brazilian viticulture?",
        "stream": False
    }
)

print(response.json()['response'])

📚 Model Details

Architecture

• Base Model: Qwen3-8B
• Fine-tuning: LoRA (Low-Rank Adaptation)
• Quantization: GGUF Q8_0 (for edge deployment)
• Context Length: 2048 tokens
• Training Data: Brazilian viticulture dataset (Embrapa + academic sources)

Training Process

1. Initial Fine-tuning: Basic adaptation to viticulture domain
2. Incremental Improvement: Progressive enhancement of Brazilian focus
3. Strategic Rebalancing: Final optimization for perfect Brazilian specialization
4. Validation: Comprehensive benchmarking against generalist models

Key Improvements in v1.3

• Prompt Rebalancing: Enhanced Brazilian terminology and professional orientation
• Ethical Guidelines: Strong emphasis on qualified Brazilian professionals
• Structure Optimization: Consistent markdown formatting for professional responses
• Edge Optimization: Adapted for low-resource environments

🎯 Capabilities

Brazilian Viticulture Expertise

• Grape Varieties: Tannat, Cabernet Sauvignon, Chardonnay, Bordô
• Regions: Serra Gaúcha, Vale dos Vinhedos, São Roque, Campanha
• Regulations: Brazilian wine laws and certifications
• Production Techniques: Traditional and modern vinification methods
• Market Knowledge: Brazilian wine industry trends and challenges

Professional Features

• Technical Accuracy: Precise terminology and current knowledge
• Ethical Orientation: Always recommends consulting qualified professionals
• Cultural Context: Deep understanding of Brazilian wine culture
• Practical Advice: Actionable recommendations for viticulturists

⚖️ License

This model uses dual licensing:

Code License (Apache 2.0)

• Applies to: Scripts, utilities, documentation
• Permissions: Free use, modification, and distribution
• File: LICENSE_CODE

Weights License (Custom)

• Applies to: Model weights and parameters
• Restrictions: Inference-only, no fine-tuning, no derivatives
• File: LICENSE_WEIGHTS

Important: The model weights are licensed under restrictive terms to protect the specialized Brazilian training investment. Commercial use for inference is permitted, but fine-tuning, model merging, or derivative creation is strictly prohibited.

📖 Citation

If you use UVIA v1.3 in your research or applications, please cite:

@misc{gandolfi2026uvia,
  title={UVIA v1.3: A Specialized Language Model for Brazilian Viticulture and Enology},
  author={Daniel Gandolfi},
  year={2026},
  publisher={Laboratório IA Uvia SLM},
  url={https://huggingface.co/uvia-br/UVIA-v1.3}
}

📄 Academic Papers

• Portuguese: UVIA v1.3 Paper (Portuguese)
• English: UVIA v1.3 Paper (English)
• Benchmarks: Complete Evaluation Report

🤝 Contributing

While the model weights have restricted licensing, we welcome contributions to:

• Documentation improvements
• Usage examples
• Brazilian viticulture data validation
• Edge computing optimizations

📞 Contact

• Author: Daniel Gandolfi
• Institution: UvIA
• Email: [email protected]
• Website: vinogandolfi.com.br
• GitHub: uvia-br

🙏 Acknowledgments

• Embrapa: Brazilian agricultural research data
• Academic Community: Brazilian enology and viticulture researchers
• Wine Industry: Brazilian viticulturists and winemakers
• Qwen Team: Base model development
• Ollama: Inference optimization

---

UVIA v1.3: Setting the standard for specialized AI in Brazilian agriculture 🇧🇷

Developed with ❤️ for the Brazilian viticulture community