Psychology-Qwen-0.5B: AI for Mental Health Support

Why This Model Exists

We have the technology to help people in their darkest moments. We have models that can understand, empathize, and provide evidence-based guidance when someone is struggling alone at 2 AM. We have the ability to make mental health support accessible to anyone with an internet connection.

But we sit and wring our hands wondering if we should while people die.

This model exists because if it helps one person, I will have done all the good I could have hoped.

For Josh. For Mark. Who had no one when they needed them.

The Reality We Face

• Therapy waitlists: 6+ months in most areas
• Crisis hotlines: Overwhelmed, understaffed, often ineffective
• Cost barriers: $200+/hour for quality mental health care
• Geographic barriers: Rural areas with zero mental health resources
• Stigma barriers: People afraid to seek help
• Time barriers: Mental health crises don't happen during business hours

Meanwhile, someone searches "I want to die" every few seconds. College students have panic attacks in dorm rooms with no one to call. Veterans wake up at 3 AM with no support system. Teenagers cut themselves because they can't articulate their pain to anyone who would understand.

This model may not be perfect, it may be clinical when we need a friend. But the alternative is silence. And that helps no one.

What This Model Does

This is Qwen 2.5-0.5B fine-tuned specifically for psychology and mental health support using PEFT/LoRA techniques. It provides:

• Immediate empathetic response when someone is in crisis
• Evidence-based coping strategies grounded in psychological principles
• Validation and understanding without judgment
• Practical techniques for anxiety, depression, trauma, grief, and more
• 24/7 availability when human support isn't accessible

Because that's how I know how to help.

Model Performance

Compared to the base Qwen 2.5-0.5B model, this fine-tuned version demonstrates:

• Significantly improved empathy: Validates emotions before offering solutions
• Practical specificity: Gives actionable advice instead of generic responses
• Psychological awareness: Understands mental health concepts and terminology
• Crisis sensitivity: Appropriate responses to expressions of self-harm or suicidal ideation

Example Comparison:

Base Model Response to "I feel like I'm not good enough at my job":

"You should set goals and strive for them, but be realistic and acknowledge your accomplishments. Overcoming your weaknesses is the most important thing."

Fine-tuned Model Response:

"When you are feeling self-doubt, it can be both a relief and a painful experience. If you find yourself struggling to see your potential in the job that you love, it's important to remember that every person has room for growth, improvement, and success. Here are some specific strategies you can try: Set realistic goals that build on your existing strengths. Be open to feedback from trusted colleagues who can offer perspective. Focus on what you're doing well and build from those successes. Recognize that everyone develops at their own pace - be patient with yourself as you grow."

Technical Details

• Base Model: Qwen/Qwen2.5-0.5B
• Fine-tuning Method: PEFT with LoRA adapters
• Training Focus: Psychology, counseling, mental health support
• Context Length: 2048 tokens
• Parameters: ~500M total, ~1.5M trainable
• Training Hardware: H100 GPU
• Training Time: ~10 minutes of focused fine-tuning, that's it...

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B")
model = PeftModel.from_pretrained(base_model, "phxdev/psychology-qwen-0.5b")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B")

# Example usage
prompt = "Human: I've been having panic attacks and I don't know what to do.\n\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=300, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Critical Disclaimers

This model is not a replacement for professional mental health care. It is designed to provide supportive guidance and evidence-based techniques as a supplement to, not substitute for, human therapeutic support.

In crisis situations:

• Call 988 (Suicide & Crisis Lifeline) in the US
• Text HOME to 741741 (Crisis Text Line)
• Call emergency services if in immediate danger

The model has been designed with safety guardrails, but AI systems can make mistakes. Always seek professional help for serious mental health concerns.

Why Open Source This?

Because mental health support shouldn't be locked behind corporate APIs or subscription services. Because researchers need to be able to study, improve, and adapt these techniques. Because the 19-year-old having their first panic attack in a college dorm room shouldn't have to wait for someone's business model to mature or tech giant to grow a conscious.

Open source saves lives.

The Bigger Picture

This model represents a proof-of-concept for what becomes possible when we fine-tune AI systems for specific human needs rather than just general capability. Mental health is just the beginning.

Imagine AI systems specifically trained for:

• Educational support for students with learning disabilities
• Grief counseling for those who've lost loved ones
• Addiction recovery support available 24/7
• Trauma processing with culturally-informed approaches
• Relationship therapy for couples who can't afford counseling

The technology exists. The methods work. The need is desperate.

What are we waiting for?

For Josh and Mark

This model exists because two people I cared about couldn't find help when they needed it most. It exists because the mental health system failed them, and technology didn't exist while they suffered alone.

If this helps one person - just one - survive a dark night when they otherwise wouldn't have, then every minute of training time, every line of code, every risk taken in deployment will have been worth it.

We build the help we wish had existed.

Contributing

If you want to improve this model, expand its capabilities, or adapt it for specific populations or use cases:

• Technical improvements: Submit PRs with better training data or techniques
• Safety enhancements: Help identify and fix potential harmful outputs
• Accessibility: Translations, cultural adaptations, specialized populations
• Research: Evaluate effectiveness, measure impact, publish findings

Contact: @phxdev

Citation

If this model helps your research or saves someone's life, consider it cited in full.

@misc{psychology-qwen-0.5b,
  author = {phxdev},
  title = {Psychology-Qwen-0.5B: Fine-tuned AI for Mental Health Support},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/phxdev/psychology-qwen-0.5b},
  note = {For Mark. For Josh. For everyone who needs someone.}
}

---

"The best time to plant a tree was 20 years ago. The second best time is now."

We can't save everyone. But we can be there for the next person who needs help.

Deploy responsibly. Help immediately. And consider donating your time to any worthy cause.