vamoruso/biscegliese-qwen2.5-gguf

🐬 Biscegliese-Qwen2.5 — Dialect LLM (GGUF)

A language model based on Qwen2.5‑7B‑Instruct, trained and optimized to understand and generate text in the Biscegliese dialect. Format: GGUF Q4_K_M, compatible with llama.cpp, LM Studio, KoboldCpp and all GGUF inference engines.

📌 Overview

Biscegliese‑Qwen2.5 is a model specialized in producing and understanding the Biscegliese dialect, with particular focus on:

• —local idiomatic expressions
• —dialect phonetics and morphology
• —IT → dialect translation
• —creative generation (poetry, dialogues, storytelling)
• —cultural and linguistic preservation

🎯 Intended Use

Recommended Uses

• —linguistic and dialectological research
• —cultural and historical applications
• —dialect content generation
• —translation and linguistic analysis

Not Recommended

• —medical, legal, financial domains
• —critical or automated decision‑making
• —sensitive or high‑risk content generation

🧠 Model Details

• —Base model: Qwen2.5‑7B‑Instruct
• —Format: GGUF
• —Quantization: Q4KM
• —Supported tasks: text‑generation, chat, translation, dialect modeling
• —Author: Vincenzo Amoruso
• —Version: 1.0

📚 Training Data

The model was trained on:

• —Biscegliese dialect corpus
• —cultural and popular texts
• —local dialogues and transcriptions
• —manually curated datasets

Possible Biases

• —non‑neutral cultural representations
• —non‑uniform phonetic variations
• —potential errors due to limited data availability

📈 Evaluation

Qualitative Evaluation

The model was tested on:

• —IT → dialect translations
• —realistic dialogue generation
• —comprehension of complex dialect sentences

Example

Input: “Come stai oggi?” Output: “Com stà ioscè?”

🔒 Safety & Ethical Considerations

The model may generate:

• —grammatical errors
• —inaccurate content
• —imperfect cultural interpretations

Do not use the model for:

• —medical diagnosis
• —legal advice
• —high‑risk decision‑making

⚙️ Environmental Impact

• —Hardware: Consumer GPU RTX5090
• —Training duration: 18 hours
• —Estimated consumption: moderate

🛠️ Inference

Esempio con llama.cpp

./llama-cli -m Qwen2.5-7B-Instruct.Q4_K_M.gguf -p "Traduci la frase in biscegliese
### Italiano:terrazzo 
### Traduzione:"