[sinuhe69]

I think the fact that they used synthetic/distilled high-quality data from GPT4-o output to train in the style of Phi models are of significance as well.