DISCLAIMER: This is an experimental testing organization. The repositories, models, and datasets hosted under this account are currently in an active testing, research, and evaluation phase.
Valleprata AI Labs 🌿🤖
Blending Authentic Umbrian Hospitality with Applied Small-Language & RAG Models
%2C%20Umbria-brightgreen?style=flat-square){kind=icon}
🏛️ About Us
Originating from Valleprata — an independent holiday home retreat situated in the heart of Umbria in Sellano (Valnerina) — Valleprata AI Labs is an experimental open initiative bridging real-world hospitality and cutting-edge generative AI.
Our mission is to prove how modern, lightweight Open-Weights models (SLMs) and Retrieval-Augmented Generation (RAG) can transform travel, hospitality, and local tourism by delivering:
- Zero-hallucination contextual guest assistants and smart concierges.
- Multilingual discovery systems for local heritage, cultural trails, and territorial points of interest.
- Privacy-conscious, edge-deployable AI tailored for regional tourism operators.
🚀 Key Research & Development Focus
1. Grounded Travel & Hospitality RAG
Building specialized adapters and datasets designed to answer inquiries accurately strictly from structured local guides (e.g., house rules, trail guides, check-in instructions, regional attractions such as the Sellano Tibetan Bridge, Rasiglia, Spoleto, and Norcia).
2. Edge & Small Language Models (SLMs)
Optimizing sub-4B models (Qwen, Llama, Gemma) via parameter-efficient fine-tuning (LoRA/QLoRA) to deliver low-latency, self-hosted AI assistants that run affordably in local or serverless environments.
3. Hyper-Local Contextual Datasets
Curation of domain-specific synthetic and real-world evaluation benchmarks focusing on:
- Itinerary generation & event summarization.
- Multi-document retrieval with distractor handling.
- Multi-turn booking & itinerary Q&A in English and Italian.
📦 Featured Repositories & Artifacts
| Model / Dataset | Description | Base Architecture |
|---|---|---|
valleprata/qwen3.5-4b-rag-specialist |
Small-footprint model tuned for high-precision RAG and unanswerable question rejection | Qwen 3.5 4B |
valleprata/hospitality-rag-eval-it-en |
Bilingual benchmark for testing extraction quality on property rules and local guides | Dataset / Synthetic |
valleprata/smart-concierge-pipeline |
Reference implementation of a hybrid vector/hybrid retriever with reranking | Python / LangChain / Haystack |
🛠️ Tech Stack & Tooling
- Foundation Models: Qwen family, Llama, Gemma
- Techniques: SFT, LoRA/QLoRA, Context Injection, Multi-document RAG
- Ecosystem:
transformers,vllm,peft,qdrant,faiss - Languages: Italian 🇮🇹, English 🇬🇧
📬 Connect & Collaborate
- 🌐 Official Website: valleprata.net
- 📍 Location: Sellano (PG), Umbria, Italy
- ✉️ Contact: info@valleprata.com
We welcome contributions, benchmark feedback, and collaborations on open-source hospitality AI!