Full training checkpoints (DeepSpeed ZeRO layout: bf16 model shards + optimizer state in global_step<N>/ + RNG/scheduler state) for resuming the run described in gr00t-n1.7-b1k-multitask.

To resume with the wensi-ai/Isaac-GR00T fork: place checkpoint-<N>/ under outputs/b1k-multitask/ and pass --resume-from-checkpoint to scripts/b1k/train_b1k.py (the trainer resumes from the highest checkpoint).