Wothmag07/counseLLM
Wothmag07/counseLLM is an 8 billion parameter conversational support model developed by Gowtham Arulmozhii, fine-tuned from Llama 3.1 8B Instruct. It utilizes a two-stage alignment pipeline, Supervised Fine-Tuning (SFT) on 36K counseling examples followed by Direct Preference Optimization (DPO) on ~2K preference-filtered pairs. This model is specifically designed for empathy-aligned conversational support, excelling in generating empathetic and supportive responses for mental health-related dialogues.
Loading preview...
CounseLLM: Empathy-Aligned Conversational Support
CounseLLM, developed by Gowtham Arulmozhii, is an 8 billion parameter model fine-tuned from Llama 3.1 8B Instruct specifically for empathetic conversational support. It employs a unique two-stage alignment pipeline to achieve its specialized capabilities.
Key Capabilities & Training
- Empathy-Aligned Responses: Achieves a GPT-4o judge score of 4.88 for Empathy and 4.71 Overall, significantly outperforming its base and SFT stages.
- Two-Stage Alignment:
- Stage 1 (SFT): Supervised Fine-Tuning on 36K multi-source counseling examples, including synthetic, clinical, and real human transcripts (e.g., MentalChat16K, empathetic_dialogues, Psych8k).
- Stage 2 (DPO): Direct Preference Optimization using ~2K preference-filtered pairs from PsychoCounsel-Preference, focusing on improving empathy, safety, relevance, and helpfulness.
- Optimized for Support: Demonstrates improved perplexity and average response length compared to its base model, indicating more coherent and detailed supportive dialogues.
Intended Use Cases
- Research and Education: Ideal for studying AI-assisted mental health support and alignment techniques in sensitive domains.
- Demonstrating Empathy: Useful for showcasing empathy-aligned language model fine-tuning.
Important Limitations
- Not for Clinical Use: This model is not a substitute for professional mental health care and is explicitly out-of-scope for clinical deployment or crisis intervention. Users should refer to professional services like the 988 Suicide & Crisis Lifeline for emergencies.
- Potential Biases: May reflect biases from its training data and could generate clinically inaccurate advice. Extensive safety review is required for any production use.