SafeTalk

SafeTalk is an AI-driven therapy simulation tool designed to train and evaluate therapists. By simulating patient personalities and providing a structured environment for therapist-patient interactions, it helps improve therapeutic skills in a risk-free setting.

The system utilizes language models like Ollama's Llama 3.2 to simulate realistic therapy conversations, and integrates a supervisor evaluation mechanism to assess therapist performance.

running make app will launch a streamlit application allowing you to chat with a patient.

Note (Mac Users)

If you choose to use the Docker ollama, run make docker-llama but be warned - on Mac there is still no option for GPU support with Docker, so it will run on the CPU and things will be very slow.

Personally, I would just run the Mac Ollama distribution and have the server running locally using your GPU.

If you're using Windows... you'll have to figure it out yourself :)

Features

• Therapy Simulation: Simulate therapy sessions between a virtual patient and a therapist.
• Customizable Patient Personas: Define different patient personalities based on predefined templates.
• Therapist Evaluation: A supervisor module evaluates the therapist's performance after each session based on set criteria.

Installation

make venv

How It Works

The pipeline runs a therapy session where a virtual patient (e.g. Mike) interacts with a therapist. The patient's personality is driven by a prompt template that simulates some mental illness, which in practice would be unknown to the therapist and part of them trying to understand the patient.

The therapist’s performance is then evaluated by a supervisor.

Key Components:

• Patient Persona: The patient has a predefined personality and emotional state (e.g., shy, self-critical, overwhelmed by stress). The patient’s responses are dynamically generated based on the therapist’s input.
• Therapist: The therapist interacts with the patient based on the conversation, adjusting their responses to foster a productive therapeutic session.
• Supervisor: The supervisor evaluates the therapist's responses based on predefined criteria (e.g., empathy, effectiveness, engagement).

Example

To run the example (using the anxiety template), just run the make command below:

make app

Explanation of the Script:

1. Load Patient Prompt: A patient persona is loaded from a template (anxiety.template) that defines Mike’s emotional state and personality traits.
2. Create LLM Model: I use OllamaLLM with the llama3.2 model to generate responses from both the patient and therapist, but you can use whatever you want.
3. Start Therapy Session: The therapy session begins with the patient and therapist interacting. The session history is tracked.
4. Supervisor Evaluation: After the session, the supervisor evaluates the therapist’s performance based on predefined criteria from the supervisor.template.

Usage

You can create new templates for different patient types or therapeutic approaches by generating new template files. I use the APA website to get case studies.

What's next?

Future Development looks like:

• Better pipelining
  • Placing prompts behind another interface