We should look into enabling best fit packing dataset curation feature. This was used by deepseek and seems like we can use our existing bin packing features to enable it so work wise it wont be too big of a lift

https://arxiv.org/pdf/2404.10830