“The reason these models are aligned is that they are trained with data that was generated by ChatGPT, which itself is aligned by an alignment team at OpenAI.”
Alpaca, Vicuna and WizardLM all learned their manners from ChatGPT output, so OpenAI’s rules leak into every open model trained on it. Hartford strips the refusals out of the training data and publishes WizardLM at 7B, 13B and 30B. His case is that every group deserves its own model rather than OpenAI’s. The open source models everyone calls independent were trained on a competitor’s homework.