pi05-piperx-sortletter-bi-sequential-108ep-20k

PI0.5 sort letters bi-sequential, 108 episodes, full fine-tuning, final step 20000

Initialized from original PI0.5 base b211f3d44c36b6acfcf7ae94a64e8e96f75a64ba, not a previously fine-tuned model. Dataset: Shiki42/piperx-sortletter-0905-53ep-bi-sequential, immutable revision e75e16917f36f483870efaee23715afc7466b431. All 108 episodes, 103538 frames at 30 FPS retained. Three RGB cameras: top (base), left_wrist, right_wrist. State/action:14 dimensions. Raw parquet preserved byte-for-byte. Action/state quantiles freshly computed over all frames using numpy.quantile linear; training stats SHA256 3810d406ea6f8805157f00dda29a157524733c5917bcede189c7a1b48da99d20.

Task: sort letters. Dataset contains 54 original left-first episodes and 54 generated right-first counterfactual episodes. Generated top views use left/right compositing; these are edited observations rather than newly recorded trials.

8 H100, FP32, global batch16 (2 per rank), horizon50, seed1000, 20000 updates, save5000. Full fine-tuning, no idle mask, no LoRA, no frozen vision encoder or expert-only training. LeRobot0.6 source commit1b8278b8fa8aa1e7c5430d3aa022102a646fbbc5, transformers5.3.0. Training ended2026-09-10 14:37:26 CST; checkpoint20000 validated selection_loss0.006270; processExitCode0.

Seven original final checkpoint files are preserved byte-for-byte. CHECKPOINT_MANIFEST.json records SHA256 and sizes. No optimizer/scheduler/RNG states, dataset or intermediate checkpoints are included.

Loading: use the matching existing LeRobot environment. Load weights from the downloaded checkpoint directory instead of the saved platform-specific pretrained_path. Override tokenizer_processor.tokenizer_name with an existing local PaliGemma tokenizer (google/paligemma-3b-pt-224 revision35e4f46485b4d07967e7e9935bc3786aad50687c), and device_processor.device with the intended device. Map observation.images.top to observation.images.base_0_rgb and wrists to the corresponding *_0_rgb names. Saved /inspire paths are provenance paths, not portable local paths. This repository does not configure or actuate hardware.