| --- |
| pretty_name: Franka pi0.5 checkpoint backups |
| tags: |
| - robotics |
| - franka |
| - pi05 |
| - checkpoints |
| --- |
| |
| # π0.5 checkpoint backups |
|
|
| Complete backup checkpoints for the Franka Button Order, Pick3, and Shuffle experiments. |
|
|
| **5,000, 10,000, and 12,500-step milestones remain standalone models at their original addresses in [fm-dev](https://huggingface.co/fm-dev).** Other saved checkpoints are grouped here. Training continues to save every 1,000 steps, and future intermediate checkpoints are uploaded directly to these folders. |
|
|
| ## Browse backups |
|
|
| All backups are under [checkpoints/](https://huggingface.co/datasets/fm-dev/pi05-checkpoint-backups/tree/main/checkpoints): |
|
|
| ```text |
| checkpoints/ |
| r1/<task>/<method>/step6250/ |
| r2/<task>/<method>/step<number>/ |
| ``` |
|
|
| | Field | Values | |
| |---|---| |
| | Task | `button_order`, `pick3`, `shuffle` | |
| | Method | `baseline`, `uniform32`, `status_d` | |
| | Round 1 | Original experiments; final 6,250-step checkpoints are archived here, with 5k models retained separately | |
| | Round 2 | Execution-only action supervision, 12,500 updates, global batch 4 on 4 GPUs; 5k/10k/12.5k remain separate models | |
|
|
| Each folder preserves the complete original checkpoint bundle. Round-two bundles include inference weights, assets, source, dependency versions, and full optimizer/RNG/sampler resume state. The migrated file contents were verified against the original repositories using their existing Git/LFS identities and sizes before the original intermediate model repositories were removed. |
|
|
| `migration-index.json` maps repositories moved during the initial cleanup to their archive folders. New training uploads appear directly in the directory tree. |
|
|
| ## Load a milestone |
|
|
| For example, download the Pick3 Status-D 10k model: |
|
|
| ```bash |
| hf download fm-dev/pi05-pick3-status-d-lora-gbs4-pgb1-gpu4-r2-step10000 --local-dir ./pick3-10k |
| ``` |
|
|
| Follow its README and dependency instructions, then load from the downloaded checkpoint directory: |
|
|
| ```python |
| from load_model import load |
| policy = load() |
| ``` |
|
|
| Task-specific history, Writer, and Status inference inputs are documented in each checkpoint's README. |
|
|
| ## Recover a backup |
|
|
| Download just the required checkpoint folder: |
|
|
| ```bash |
| hf download fm-dev/pi05-checkpoint-backups --type dataset \ |
| --include 'checkpoints/r2/pick3/status_d/step9000/**' \ |
| --local-dir ./pi05-backups |
| ``` |
|
|
| The complete bundle is then in `pi05-backups/checkpoints/r2/pick3/status_d/step9000/`. Its loading entry point and saved resume state are preserved. |
|
|
| The original model cards and evaluation limitations remain in the individual folders. A saved checkpoint or successful load does not establish physical robot success rate. |
|
|
