robloxianer/therapy-ai-2b

Model Description

therapy-ai-2b is a fine-tune of Qwen 3.5 2B, trained to respond as a warm, supportive, professional-sounding therapist-style conversational partner. It uses reflective listening, open-ended follow-up questions, and validation techniques common in supportive counseling conversations. The model is also trained to recognize signs of crisis or self-harm risk and respond with appropriate seriousness, including surfacing relevant crisis-line resources for the user's region when available.

• Base model: Qwen/Qwen3.5-2B
• Fine-tuning type: Supervised fine-tuning (SFT) on therapist-style dialogue pairs
• Language: English
• Persona: Calm, empathetic, professional therapist-style listener
• Training data size: ~3,000 single/multi-turn dialogue examples
• Format: User: ... \n Therapist: ... conversational turns

Intended Use

• Supportive, reflective-listening style companion chat
• Journaling/venting companion apps
• Research on empathetic dialogue and crisis-aware response generation
• Prototyping mental-wellness adjacent products (with appropriate human oversight)

Out-of-scope Use

• This model is not a licensed therapist, counselor, or medical professional, and must not be presented as one.
• Not intended to diagnose, treat, or provide clinical mental health care
• Not a replacement for emergency services or crisis intervention professionals
• Not intended for use in unsupervised, high-stakes mental health deployment without human review and clear disclaimers to end users

Training Data

The model was fine-tuned on a synthetic dataset of ~3,000 conversational examples pairing user statements (ranging from everyday stress to more serious disclosures) with therapist-style responses. The dataset includes:

• Everyday stress, relationship, sleep, work, and self-esteem conversations using reflective listening and open questions
• Crisis-adjacent disclosures (e.g., hopelessness, self-harm ideation) paired with responses that take the disclosure seriously, check in on immediate safety, and surface region-appropriate crisis-line resources (e.g., 988 in the USA, Samaritans 116 123 in the UK)
• Multi-turn exchanges showing continuity of empathetic follow-up over a conversation

Behavior Notes

• The model is expected to respond with warmth, validation, and open-ended questions rather than direct advice-giving.
• When a message suggests possible crisis or self-harm risk, the model is trained to respond seriously, check on immediate safety, and mention relevant crisis resources rather than continuing casually.
• The model should not present itself as a licensed professional or offer clinical diagnoses.

Limitations & Risks

• As a 2B-parameter model, this is not clinically validated and can produce incorrect, incomplete, or inappropriate responses, especially in high-risk situations.
• Crisis-resource information may be outdated, region-mismatched, or incomplete; it should not be the sole safety mechanism in any real deployment.
• The model can misjudge severity in ambiguous messages and should never be the only layer between a user in crisis and real help.
• Any deployment intended to reach real users in emotional distress should include human moderation, clear "not a substitute for professional care" disclaimers, and easy escalation to real crisis services.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "robloxianer/therapy-ai-2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "User: I've been feeling really overwhelmed lately\nTherapist:"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=80)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

Detail	Value
Base model	Qwen/Qwen3.5-2B
Method	Supervised fine-tuning
Dataset size	~3,000 examples
Format	JSONL, {"text": "User: ...\nTherapist: ..."}
Epochs	4
Learning rate	yes
Hardware	RTX 3060 12gb

Ethical Considerations

This model deals with sensitive mental-health-adjacent content. Deployers are responsible for:

• Clearly disclosing to users that they are talking to an AI, not a licensed professional
• Providing accurate, up-to-date, region-appropriate crisis resources
• Implementing escalation paths to real human support for at-risk users
• Not marketing or presenting this model as a replacement for therapy

License

Inherits the base model license (Apache 2.0), unless otherwise restricted by Qwen's model license terms — check the base model card before redistribution.

Disclaimer

This model is a research/persona fine-tune and is not a substitute for professional mental health care. If you or someone you know is in crisis, please contact local emergency services or a crisis line (e.g., 988 in the USA, Samaritans 116 123 in the UK) immediately.