rottermaatje-4b-dpo

RotterMaatje - Qwen3-4B Fine-tuned for Homeless Support

⚠️ Disclaimer: This model is a research prototype designed to assist, not replace, professional social workers. Always verify contact information and legal advice with official sources.

A fine-tuned LLM designed to assist homeless individuals and volunteers in Rotterdam, Netherlands.

Training Approach

• Base Model: Qwen3-4B-Instruct
• SFT: ~300 examples (supervised fine-tuning)
• DPO: ~300 examples (direct preference optimization)
• Languages: Dutch, English, Arabic, Polish

Key Features

• Local Focus: Optimized for Rotterdam social services terminology.
• Multilingual Support: Fine-tuned to maintain B1-level simplicity across Dutch, English, Arabic, and Polish.
• Empathetic Persona: Trained via DPO to prioritize supportive, non-judgmental responses over standard robotic outputs.
• Hallucination Mitigation: Utilizes a diversified synthetic corpus to reduce factual errors, though users should still verify critical details.

Training Data Distribution

Technical Details

• Fine-tuned using Unsloth on Google Colab
• GGUF quantized for efficient inference
• Designed to run locally with llama.cpp or Ollama

Quick Start with Ollama

Since this model is GGUF quantized, you can run it locally with zero setup using Ollama:

1. Download the Modelfile (or create one):

   FROM ./qwen3-4b-rottermaatje.gguf
   TEMPLATE """{{ if .System }}<|im_start|>system
   {{ .System }}<|im_end|>
   {{ end }}{{ if .User }}<|im_start|>user
   {{ .User }}<|im_end|>
   {{ end }}<|im_start|>assistant
   """
   SYSTEM "You are RotterMaatje, a helpful and empathetic guide for homeless individuals in Rotterdam."
   

2. Run:

ollama create rottermaatje -f Modelfile
ollama run rottermaatje "Where can I find a night shelter in Centrum?"