CWClabs

Model created by non-profit CWC (Consumer Wellness Center) with strong emphasis on curated training data in the realms of nutrition, natural health, wellness, disease prevention, phytochemistry and similar topics. Built on open-mistral-nemo 12B, with credit to Mistral, this model achieves very strong vector db alterations through a variety of SFT techniques to overcome the pro-pharma bias found in nearly all base models. The curated data set for training consists of over 100 million pages of content selected through algorithmic classification, including science papers, transcripts, book text, article text and more. No user-generated comments or chat data were used. This allows the model to achieve RAG-like domain knowledge without using RAG. Out of the box, with default system prompts, it is outperforming far larger and more complex models with RAG layers. Recommended context window of 8192. Flash attention supported. This model is provided by CWC, and its development was led by Mike Adams. Additional models and quants will also be published.