Wothmag07/counseLLM

Overview

Wothmag07/counseLLM is an empathy-aligned conversational support model developed by Gowtham Arulmozhii, built upon the Llama 3.1 8B Instruct base model. It was fine-tuned using a two-stage alignment pipeline: Supervised Fine-Tuning (SFT) on 36,000 counseling examples, followed by Direct Preference Optimization (DPO) on approximately 2,000 preference-filtered pairs. This process significantly improved its performance in empathy, safety, relevance, and helpfulness, as measured by LLM-as-Judge evaluations using GPT-4o.

Key Capabilities

• Empathy-Aligned Responses: Achieves a 4.88 empathy score (GPT-4o judge) through specialized DPO training.
• Conversational Support: Designed to provide supportive, empathetic guidance, acknowledging feelings and exploring situations with open-ended questions.
• Safety and Relevance: Demonstrates high scores in safety (4.60) and relevance (4.88) for mental health contexts.
• Specialized Training Data: Utilizes diverse counseling datasets including synthetic, clinical, and real human multi-turn dialogues.

Good For

• AI-assisted Mental Health Research: Ideal for studying and developing AI applications in mental health support.
• Alignment Technique Research: Useful for exploring SFT and DPO methods in sensitive domains.
• Demonstrating Empathy: Showcasing fine-tuning for empathetic language generation.

Limitations

It is crucial to note that counseLLM is an AI research project and not a substitute for professional mental health care or clinical deployment. It should not be used for crisis intervention or as a replacement for licensed therapists due to potential biases, generic responses, or clinically inaccurate advice.