Wothmag07/counseLLM
Wothmag07/counseLLM is an 8 billion parameter causal language model developed by Gowtham Arulmozhii, fine-tuned from Llama 3.1 8B Instruct with a 32K context length. It is specifically aligned for empathetic conversational support, utilizing a two-stage Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) pipeline on diverse counseling datasets. This model excels at providing supportive, empathetic guidance for research and educational purposes in AI-assisted mental health support.
Loading preview...
CounseLLM: Empathy-Aligned Conversational Support
CounseLLM is an 8 billion parameter causal language model developed by Gowtham Arulmozhii, built upon Meta's Llama 3.1 8B Instruct. Its core distinction lies in its specialized two-stage alignment pipeline, designed to enhance empathetic conversational capabilities for mental health support.
Key Capabilities & Training
- Empathy-Aligned: Fine-tuned using a two-stage process: Supervised Fine-Tuning (SFT) on 36,000 multi-source counseling examples, followed by Direct Preference Optimization (DPO) on approximately 2,000 preference-filtered pairs.
- Optimized for Support: Achieves a perplexity of 3.13 and significantly higher scores in empathy, safety, relevance, and helpfulness (4.88, 4.60, 4.88, 4.48 respectively on a 1-5 GPT-4o judge scale) compared to its base and SFT versions.
- Robust Training Data: SFT utilized diverse datasets including MentalChat16K, empathetic_dialogues, Psych8k, counsel-chat, and ESConv, while DPO leveraged the PsychoCounsel-Preference dataset.
Intended Use Cases
- Research & Education: Ideal for studying AI-assisted mental health support and alignment techniques in sensitive domains.
- Demonstration: Useful for showcasing empathy-aligned language model fine-tuning.
Important Limitations
It is crucial to note that CounseLLM is not a substitute for professional mental health care and is explicitly not intended for clinical deployment, crisis intervention, or as a replacement for therapy. Users should be aware of potential biases, generic responses, and the risk of clinically inaccurate advice.