Modelpublic

Shiki42/s015-scan-object-ctr50-withmask-pi05-step20000

sourceHugging Faceupdated 4d agoView on Hugging Face

0likes

Model Card

E746 Scan Object PI0.5 checkpoint — step 20000

Inference checkpoint from the S015 CTR50 with IdleMask ablation. It contains the OpenPI parameter tree and normalization assets. Optimizer and training loader state are excluded.

Training provenance

—Experiment / Run: E746 / E746-R001
—Policy: PI0.5 JAX LoRA; checkpoint step 20000 of 30,000
—Initialization: XinY0201/openpi-pi05-base-jax at 5e62884fcf8cb8f9fc693c9163ea18d3e3739658
—Seed: 87431; batch size 16; gradient accumulation 1
—Dataset: `Shiki42/ctr-scan-object-uniform50-20260917` at revision a34f2a32d3738e3607f226a777b41d42d0522908
—Data: Uniform50, 50 episodes / 14,163 frames
—IdleMask: enabled as the registered S015 ablation; the action loss consumes observation.arm_active_mask; standard temporal padding mask is retained
—Normalization: exact global NumPy quantiles at 1% and 99%, linear interpolation; artifact SHA-256 e08ec02f286016017abf90e5be9f11e7507bd4b93a1b68473ac22dd7c3fb62b8
—Normalization mask rule: statistics use every unpadded action-window element with unit weight; the training loss still consumes IdleMask
—CTR commit: f604dc908e21458d71c3ffb6baed24c523188a2e
—OpenPI commit: e9ba7b7732a3e66e4bd87d6d3429f7cef6352ead
—Runtime: 2026-08-22.2; manifest SHA-256 06e035fbc62dd7f32f01a2f5ad536f01bd5af6caae3239698dbb8c07ed503332
—Checkpoint verification: fresh-process reload and finite-parameter checks passed for steps 10k, 20k, and 30k
—Evaluation: pending; this card makes no success-rate claim

Files

—params/: OpenPI model parameters
—assets/: normalization assets
—resolved_config.json: resolved inference configuration with machine-local paths removed
—training-provenance.json: pinned training, dataset, normalization, and runtime identities