Ollama - testing best way of calling models, and comparing models - June 28, 2024

Ollama - testing best way of calling models, and comparing models - June 28, 2024 Presented by Vitalii Stinskii Slides - https://github.com/lselector/seminar/tree/master/2024 --------- Contents of today's video: Using Ollama: - Selecting Best Way of Calling LLMs - Selecting Best Model to Use We have: - tested several LLM calling methods using Ollama locally - tested several relatively small models (qwen2, mistral, llama3) - used only "instruct" versions of these models - compared calling methods and models Results: - Best calling method is to use simple "instruct" prompt - Best model - qwen2:7b-instruct-q4_0 - Generally Qwen2 is better than Mistral, which is in turn better than Llama3 - Surprisingly Qwen2 4bit is better than Qwen 8bit (in our tests) More links: - https://vitalii-stinskii.com - https://github.com/vstinskii/ai_chatbots - AI Chatbots - https://EAIS.ai - Enterprise AI Solutions