RoboCasa TTT experimental checkpoints
Transfer complete: all 386 planned model/configuration files were verified against remote file sizes on 2026-09-21.
Migration documentation and TODO: https://github.com/yyoraa/ttt_memory/tree/main/experiments/2026-09-20-handoff (private GitHub repository; access required). Raw results: https://huggingface.co/datasets/yyoraa/ttt-experiment-results
Priority resume checkpoint: ttt/navigation-stage2-20k-4gpu-20260919/checkpoint-650.pt. This includes the complete model, AdamW state, Python/NumPy/PyTorch/CUDA and per-rank random states. Training is paused. Logged steps651–672 were not checkpointed and must be recomputed.
Navigation changed from four ranks to two after step400 while retaining global batch4; it is not a bitwise continuation of the original four-rank sampling stream. The training target is1000 updates, with the original20000-step learning-rate schedule retained. See GitHub TODO for all remaining experiments and limitations.
Weights are experimental derivatives of their respective upstream models. Upstream licensing/usage terms continue to apply; this transfer does not grant a new blanket license. Only load pickle-based .pt checkpoints from sources you trust.