Nicely written, thanks for posting!

I was reminded about recent LLM wins coming from training data improvements (eg. fineweb)