YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
BEHAVIOR pi0.5 all-100 preparation
Public experiment storage under ryanjin333. Production training is active; no evaluated all-100 success rate is available yet.
Official pi0.5 base will be fine-tuned on all100 tasks using RGB, state and actions. One preemptible Nebius RTX PRO6000 trains; all evaluation and replay use the existing Linux machine. Six-update full-model qualification passed after fixing the video-decoder cache. Measured steady throughput was19.9245seconds per effective-batch256 update, supporting an11k-update candidate with saves every40updates within a72-hour cap; measured checkpoint overhead gives an estimated68.8hours before startup and interruptions. This is not a success-rate prediction. Production is running on the single GPU under the fixed72-hour deadline. Update55 was observed with finite loss; checkpoints1and40 finalized and their public recovery objects were independently verified without authentication. Milestone evaluation remains pending.
Full checkpoint6 was uploaded to the public recovery bucket and independently downloaded with byte integrity verified. A fresh GPU process restored that downloaded state and completed update7 with finite loss. The integrated VM interruption drill has also restarted automatically and restored checkpoint6; it completed updates7and8, finalized checkpoint8 and stopped the VM. Automatic HF fallback and subsequent local byte verification also passed. The exported policy passed two synthetic-observation inference sessions on the local RTX3090. One real picking_up_trash simulator episode completed7902steps at the official horizon with success=false and Q=0. This six-update policy qualifies the pipeline, not performance.
Storage
- This model repository: versioned plans, source changes, manifests, qualification evidence and milestone policies.
- Public bucket
ryanjin333/behavior-pi05-all100-recovery: rotating full training-state backups; retain the latest two verified recovery points. - Dataset repository behavior-pi05-all100-rollouts: rollout videos and per-episode metrics.
The public bucket tiny-model drill restored optimizer/model/update/EMA state and normalization and matched the next uninterrupted update. This is transport qualification, not a trained robot policy. Full pi0.5 remote restore and simulator admission passed; production training is active; milestone evaluation and broad all100 performance remain pending.
See planning/launch-plan.md, planning/run-contract.json and planning/public-bucket-recovery-evidence.json for the current preparation state. Credentials are never published.