🍷 Piemontese-Gioan (LoRA Adapter)
English
Overview
Piemontese-Gioan is a LoRA adapter for Qwen2.5-1.5B-Instruct, developed by YecoAI Research. It is fine-tuned to converse in the Piedmontese dialect with a cordial, ironic, and natural tone, reflecting the most authentic "Sabaudo" (Savoyard) spirit.
The character is Gioan, a gentleman from Piedmont who chats with the typical spirit of someone in a traditional "piola" (tavern) in Turin, between a glass of wine and a talk about the weather.
🎯 Intended Use
- Primary Use Case: Conversational AI in Piedmontese dialect for entertainment and cultural preservation.
- Target Audience: Researchers of low-resource languages and dialect enthusiasts.
- Out of Scope: Not suitable for formal legal or administrative translations.
⚙️ Technical Specifications
- Framework: PEFT (Parameter-Efficient Fine-Tuning) via LoRA.
- Base Model: Qwen/Qwen2.5-1.5B-Instruct
- Training Method: Supervised Fine-Tuning (SFT) on curated dialectal conversational pairs.
- Optimizer: AdamW with linear schedule.
⚠️ Limitations & Biases
- Linguistic Nuances: Piedmontese has deep regional variations; Gioan is optimized for a natural, common version.
- Model Size: Being a 1.5B model, it might occasionally struggle with very complex logical reasoning in dialect.
- Safety: Inherits the safety guardrails and alignment of the base Qwen model.
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch
base_model = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "YecoAI/piemontese-gioan"
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype=torch.float32, device_map="cpu")
model = PeftModel.from_pretrained(model, adapter_id)
Italiano
Descrizione
Piemontese-Gioan è un LoRA adapter per Qwen2.5-1.5B-Instruct, sviluppato dalla divisione Ricerca di YecoAI. È stato addestrato per conversare in dialetto piemontese in modo cordiale, ironico e naturale.
💬 Esempi di Chat Divertenti / Fun Examples
User: "Ciao Gioan, come stai oggi?" Gioan: "Cerea bèl fieul! I stago bin, i soma sì ch'i spetoma che la jurnà a finissa për andé a beive un bicer. E ti lòn ch'it conte? Spero ch'it l'abie nen trop da fé, neh?" (Translation: "Greetings, my dear boy! I'm doing well, just waiting for the day to end so I can go have a glass. And you, what's your story? I hope you don't have too much to do, eh?")
User: "Gioan, ma è meglio il vino rosso o il bianco?" Gioan: "Ma che domanda it mi fase? 'O vin ross a l'é la salute, a l'é 'l sang de la terra! Un bel bicer de Barbera e it vedi che 'l mond a smija sùbit pì bel. 'O bianch... va bin per le fumele o per la pessera, ma per noi autri a-i vòl ròba fòrta!" (Translation: "What a question! Red wine is health, it's the blood of the earth! A nice glass of Barbera and you'll see the world looks beautiful right away. White... it's fine for the ladies or for fish, but for us, we need the strong stuff!")
🔬 About YecoAI
YecoAI is a research startup (owned by highmark.it) focused on researching the next layer of intelligence. We apply rigorous, reproducible methodologies to turn frontier ideas into dependable systems.
- Research-first: Rigorous evaluation and transparent trade-offs.
- Privacy-first: EU-ready compliance and consent-aware data strategies.
- Cognitive Architectures: Developing layers like Anti-Loop for stable agent autonomy.
Learn more at yeco.ai. "Researching the next layer of intelligence."