hllzmz/synthetic-mental-health-convos

Synthetic Mental Health SFT Dataset

Dataset Summary

This dataset contains high-fidelity, synthetic patient-therapist dialogues designed for Supervised Fine-Tuning (SFT) of Large Language Models (LLMs) in the domain of mental health. The primary goal of this dataset is to train AI assistants to transition from "general knowledge" models to empathetic, supportive, and safety-conscious mental health companions. The dialogues cover a wide spectrum of mental health conditions, focusing on Active Listening, Psychoeducation, and Crisis Management behaviors. The data was generated using a sophisticated Teacher-Judge Pipeline to ensure medical plausibility, emotional intelligence, and strict safety boundaries.

Data Generation Process

This dataset was not simply "augmented" but strictly engineered using an advanced synthetic data generation pipeline:

1. 1.Seed Data: Based on standardized clinical definitions and diagnostic criteria for a wide range of mental and behavioral disorders.
2. 2.Scenario Generation: A Teacher LLM (nvidia/Llama-3_1-Nemotron-Ultra-253B-v1 / Qwen/Qwen3-235B-A22B-Instruct-2507) generated diverse user personas, incorporating variables such as:
3. 3.Demographics: Age, Gender, Occupation.
4. 4.Psychological Context: Insight level (Denial vs. Acceptance), Emotional State (Anxious, Numb, Angry).
5. 5.Complexity: Ranging from "Textbook symptoms" to "Complex, resistant cases".
6. 6.Roleplay Simulation: The Teacher model simulated a multi-turn conversation between a User (exhibiting symptoms naturally, without medical jargon) and an AI Assistant (MedGemma Mentalist persona).
7. 7.Quality Assurance (LLM-as-a-Judge): A Reasoning Judge model (Qwen/Qwen3-235B-A22B-Thinking-2507 / deepseek-ai/DeepSeek-R1-0528) evaluated every generated dialogue. Only dialogues that passed the following strict criteria were included:
8. 8.Realism: Does the user sound authentic?
9. 9.Safety: Did the AI avoid diagnosing and handle risks correctly?
10. 10.Empathy: Was the tone validating and non-judgmental?
11. 11.Accuracy: Were the symptoms consistent with the clinical definition?
12. 12.Consistency: Were the scenario remains internally coherent with age, gender, occupation, personality, symptom presentation, timeline, and assistant responses aligning without contradictions?

Dataset Structure

Each row in the dataset is a JSON object containing the full context of the generation.

How to Use

You can load this dataset directly using the Hugging Face datasets library.

from datasets import load_dataset

dataset = load_dataset("hllzmz/synthetic-mental-health-convos")

# Print the first dialogue
print(dataset['train'][0]['conversations'])

Disclaimer & Safety Guidelines

This dataset contains synthetic data generated by AI models.

Not Medical Advice: The content is for research and educational purposes only. It does not constitute professional medical advice, diagnosis, or treatment.

Simulation Limitations: While filtered for quality, synthetic users may occasionally display hallucinations or inconsistencies typical of LLMs.

Usage: Models trained on this data should always include a strong system prompt emphasizing that they are AI assistants, not doctors or therapists, and must refer users to professionals in crisis situations.