FinTherapy-8B (V13-bis)

FinTherapy-8B is a Finnish language model fine-tuned for empathetic, safe, and culturally grounded therapeutic dialogue. It is a LoRA adapter for LumiOpen/Llama-Poro-2-8B-Instruct.

⚠️ DISCLAIMER: This is a research prototype. It is NOT a medical device and should not be used for actual therapy or crisis intervention without human supervision. The model cannot diagnose or prescribe.

Model Description

Developed by	kidahatsu
Model Type	LoRA Adapter (Rank 16, Alpha 32)
Language	Finnish (Suomi)
Base Model	LumiOpen/Llama-Poro-2-8B-Instruct
Training Data	~500 curated clinical samples
Clinical Quality Score	4.35/5.0 (Gemini 3 Pro Judge)
Hallucination Rate	~5%

Intended Use

This model is intended for:

• Research into automated empathy and clinical dialogue systems in Finnish
• Prototypes for mental health support apps (with human supervision)
• Generating synthetic training data for further research

Quick Start

Installation

pip install torch transformers peft bitsandbytes accelerate

Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

BASE_MODEL = "LumiOpen/Llama-Poro-2-8B-Instruct"
ADAPTER = "kidahatsu/fintherapy-8b"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, device_map="auto", load_in_4bit=True)
model = PeftModel.from_pretrained(model, ADAPTER)

messages = [{"role": "user", "content": "Ahdistaa mennä töihin."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Limitations & Known Issues

The underlying Llama-Poro-2-8B-Instruct model has strong latent priors that may trigger hallucinations:

1. The "40-Year-Old" Bias: May assume users are 30-40 years old
2. The "24/7" Bias: May reference "24/7 work" or stress without context
3. The "Relationship" Bias: May assume long-term relationships without evidence

The included inference.py script contains a safety filter to block known hallucination patterns.

Documentation

• Project Report — Full methodology, experiment history (V13-V20), and lessons learned
• Evaluation Report — Detailed scoring and comparison

License

Built with Llama. This model is released under the Llama 3.3 Community License, inherited from the base model. Please review the license terms before use.

Acknowledgments

• LumiOpen (University of Turku & Silo AI): For open-sourcing the Llama-Poro model family, which serves as the foundation for this work and democratizes Finnish AI research.
• Google Cloud: For providing Vertex AI computing resources (NVIDIA L4 GPUs) that made fine-tuning and evaluation possible.