CounselorChat (C2) — A Mental Health Conversational Assistant

Overview

Features

• Base Model: Llama 3.1 8B.
• Training Process:
  • Datasets: Fine-tuned on a combination of:
    • CalebE/new_mental_health_conversations_all1
    • Amod/mental_health_counseling_conversations
  • Method: One epoch of training using Unsloth at 16-bit precision.
  • Data Cleanup: Deduplication and cleaning were performed to enhance the quality of training data.

Deployment Options

1. GGUF Quantized Model: Published as an 8-bit GGUF quantized model. Other quantization levels are available upon request.
2. Modelfile for Ollama: A pre-configured Modelfile for deployment on Ollama is also included, with a default context length of 8192 tokens.
   • To deploy, update the PATH_TO_MODEL at the beginning of the Modelfile to point to the GGUF file location.
   • Then, run the following CLI command from the same directory as the Modelfile:
     ollama create C2_llama3.1_8B_q8x8K --file OllamaModelfile_C2_llama31_8B_Q8x8K
3. Other Deployment Options: For alternative deployment setups or platforms, it is recommended to set the SYSTEM prompt to:

   You are an AI counselor designed to provide compassionate, evidence-based mental health support. Offer helpful, non-judgmental guidance focusing on validated techniques. Avoid diagnosing conditions, or speculative or unsafe advice.

Intended Use

• Offering general advice for stress management, mindfulness, and coping strategies.
• Engaging in non-clinical conversations.

Limitations

• The model is not equipped to handle emergencies or complex mental health cases.
• Responses are generated based on the training data and may lack the nuance of human judgment.
• While efforts have been made to clean and deduplicate the datasets, unintended biases or inaccuracies may persist.