Llama-Poro-2-8B-Instruct-Viaraleuka
This model is a fine-tuned version of LumiOpen/Llama-Poro-2-8B-Instruct specifically trained to provide technical support for Home Assistant in the Savonian dialect (Savo).
Model Description
The goal of this model is to create a leathery, "kiero" (twisted/witty) persona that understands the deep nuances of Finnish regional dialects. Instead of acting like a rigid technical support agent, this model excels at providing daily announcements, status reports, and casual observations about the home environment in a way that feels human and regionally authentic.
Key Objectives:
- Low-Resource Efficiency: Optimized to run locally on hardware with limited power, such as the Beelink SER5, ensuring privacy and independence from cloud services.
- Narrative Reporting: Designed to transform raw Home Assistant data (sensor logs, camera alerts, weather updates) into colorful, Savonian-style situation reports and summaries.
- Dialect Preservation: Captures the "kiertelevä" (indirect) yet insightful communication style typical of the Savo region, making home automation feel less like a machine and more like... a local character :D
Further Training Objectives:
- More intense pronouncing
- More region specific vocabulary
- Less hallucination in numeric output
Intended Use
- Status Reports: Giving leathery updates on home metrics like temperature or energy usage. ("Sauna on jo satakymmenen astetta! Jos et oo tekemässä palvaushommia, niin ois varmaan aika vääntää kiuas pienemmälle")
- Summaries: Summaries about daily trends like weather or energy price. ("Aamulla on vielä viileetä eli viisi astetta, mutta aurinko paistaa. Ota aurinkolasit matkaan, ettet joutu siristelemään, ja pukeudu kerroksittain. Iltapäivällä kymmenessä asteessa alkaa jo tulla hiki jos on liikaa päällä")
Examples of the Instructions:
Instruction: Varoita, että olohuoneen kosteusprosentti on noussut yli 80%.
Response: Nyt on olohuoneessa ilma niin paksua, että siellä voisi vaikka uija. Kosteusmittari näyttää jo yli kaheksankymmentä prosenttia. Pitäsköhä laittaa ilmanvaihtoa isommalle tai aukasta räppänä?
Instruction: Tee lyhyt yhteenveto tulevasta säästä. Sääennuste: 2026-03-29T19:00:00+03:00: cloudy, lämpötila 4.3 °C, tuuli 9.36 m/s, sademäärä 0.02 mm.
Response: Iltaseitsemältä on harmaata ja pilvistä. Lämpötila on neljä pilkku kolme astetta ja tuuli puhaltaa lähes kymmentä metriä sekunnissa. Ihan vähän saattaa ripsiä vettäkin, joten ota huppu mukaan iltalenkille.
Instruction: Ilmoita Pekan kalenterimerkinnästä: Alkaa 17:00 - Taloyhtiön kokous. Neuvo asenteesta.
Response: Pekalla on illalla viideltä se taloyhtiön kokous. Vedä syvään henkeä ja ota kärsivällisyyttä matkaan, siellä varmaan taas vängätään turhanpäiväisistä asioista tunti sotisopalla. Varaa mukaan kynä ja ehkä vähän huumorintajua.
Training Details
The model was trained on a custom-curated dataset of synthetic 850 conversational rows generated with Gemini 3 Flash, covering:
- Savonian linguistic properties, talking style and proverbs
- Smart home specific instructions
Training Procedure
- Tooling: Trained using Unsloth and Hugging Face
SFTTrainer. - Hardware: Kaggle T4 GPU (for training).
- Epochs/Steps: 300 steps (~2.8 epochs).
- Learning Rate: 2e-4.
- LoRA Configuration: rank=16, alpha=32.
- Quantization: 4-bit NormalFloat (BNB).
Evaluation Results (Training Logs)
The first iteration of model reached a final training loss of approximately 0.14 - 0.20, indicating high convergence on the dialect style while maintaining technical accuracy.
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
- Downloads last month
- 7
Model tree for Jarauvi/Llama-Poro-2-8B-Instruct-Viaraleuka
Base model
LumiOpen/Llama-Poro-2-8B-Instruct