ajibawa-2023/carl-7b

What is Carl-7b?

ajibawa-2023/carl-7b is a 7 billion parameter language model developed by ajibawa-2023, specifically designed to function as a "Therapist AI." This model aims to provide conversational support for individuals experiencing stress, offering quick responses and assistance in calming down. It was trained using the FastChat codebase on Azure with 4 x A100 80GB GPUs for 3 epochs, a process that took 22 hours.

Key Capabilities & Training

• Therapeutic Conversation: Carl-7b is trained on a large synthetic dataset comprising over 100,000 sets of conversations, each containing 10-15 exchanges between Carl and a client. This synthetic data approach addresses the scarcity of publicly available, directly applicable therapy conversation data for LLMs.
• Stress Support: The model is designed to answer questions and help users calm down, acting as a conversational aid in stressful situations.
• Research Focus: It is presented as a research experiment exploring the potential of AI in therapeutic contexts, emphasizing that it is not a replacement for professional medical or psychological help.

Performance Benchmarks

Carl-7b's performance on the Open LLM Leaderboard shows an average score of 40.87. Notable scores include:

• ARC (25-shot): 53.5
• HellaSwag (10-shot): 78.29
• MMLU (5-shot): 33.96
• TruthfulQA (0-shot): 40.29

Important Considerations

It is crucial to understand that Carl-7b is a research experiment and should not be used as a substitute for a professional therapist or doctor. Users experiencing stress or difficult times are strongly advised to seek professional help or confide in friends and family.