Analysis
This article explores the fascinating world of "Base Models" in the realm of 生成AI, showcasing what Large Language Models look like before alignment training. The author uses Ollama to interact with a Mistral 7B Base Model, highlighting the differences between unaligned and aligned models. It's a fantastic look at the fundamental building blocks of modern AI.
Key Takeaways
Reference / Citation
View Original"The Base Model doesn't see 'hello' as a greeting; it's just a Japanese token string. The result of probabilistically predicting 'text that is likely to follow this' landed on a Japanese anime blog, which it likely saw in its training data."