Building an LLM from scratch is huge effort. That is why I assume it hasn't happened too often. If we exclude toy and research models and restrict the scope to language models and multimodal modals with a language component, do we know the number of models trained from scratch so far?