Blend for Super 3 and Training Questions

[berserkr]

Hi, I noticed that the blend both raw and tiny are empty for super_v3. Will you be sharing the blends with proportions at some point? Also, I noticed that in the paper there are 2 Phases mentioned, Phase 1 at 256k and Phase 2 at 512k, do you have those samplings as well? Finally, the chat template strips reasoning data from multi-turn samples right? How is this controlled during training? Thank you!