FasterWAM checkpoints

Checkpoint collection for four real-world tasks and four model families.

Layout: <task|joint>/<model>/<run_id>/checkpoints/.

• Tasks: fruit_pick, dish_utensil, place_cube, push_t.
• Models: fastwam, fasterwam (formerly FlashWAM), lingbot-va, pi0.5.
• joint is a top-level folder alongside the four tasks, for models trained on all tasks.
• Task-specific runs sit directly under <task>/<model>/.

Each run includes its original checkpoint format and inference/training metadata. Reserved folders contain a README; inspect run folders for available weights.

See CHECKPOINT_UPLOADS.md for the upload contract and commands, checkpoint_layout.json for the path schema, and hf_checkpoint_imports.json for pinned source revisions and file hashes. Imported run READMEs retain their original training and evaluation limitations; these historical runs are not asserted to use the current benchmark split.

Current corrected-data joint checkpoints — September 21, 2026

All four runs use the four tasks together and the verified corrected place_cube dataset revision 4ad42310b4408ff08370124e680e3a20ad267b1e, with 360 train / 40 held-out episodes and global batch 32 on 8 H200 GPUs.

Model	Current run	Published epochs / steps	Training outcome
FasterWAM	4414	10, 20, 30	Time limit after epoch 30 was saved
FastWAM	4415	10, 20	Time limit during epoch 26; no epoch-30 checkpoint exists
pi0.5	4441	30, 40, 50	Completed 50 epochs; final checkpoint published
LingBot-VA	4469	Steps 1,200 / 1,500 / 3,000	Completed 3,000 updates; final checkpoint published

Native WAM steps for epochs 10/20/30 are 29,310 / 58,620 / 87,930. The published pi0.5 epochs 30/40/50 use zero-based labels 87,929 / 117,239 / 146,549 (87,930 / 117,240 / 146,550 completed updates). These are inference bundles; optimizer/resume state is excluded.

Previous incorrect-data runs are preserved under each model's old/: FasterWAM archive, FastWAM archive. No prior joint pi0.5 run existed in this repository. Validation metrics are model-specific and are not closed-loop robot success measurements.

Imported archives

Source	Run folder
SleepMastger/pusht-fastwam	push_t/fastwam/pusht-fastwam-222aa512
SleepMastger/dish-utensil-lingbot	dish_utensil/lingbot-va/dish-utensil-lingbot-ca6e4fe0
SleepMastger/fruit-picking-lingbot	fruit_pick/lingbot-va/fruit-picking-lingbot-0613ebe6
SleepMastger/fruit-picking-fastwam	fruit_pick/fastwam/fruit-picking-fastwam-ebb79ac0
SleepMastger/dish-utensil-fastwam-v2	dish_utensil/fastwam/dish-utensil-fastwam-v2-601d741a

Completed local training runs

Run	Native checkpoints	Final	Contents
fruit_pick/pi0.5/train90_job3977	5000, 10000, 15000, 16769	16769	params and normalization (inference)
place_cube/fastwam/train90_scratch_job4135	2620, 7860, 13100	13100	Three inference checkpoints and validation results
joint/fastwam/train90_joint_job4146	29310, 58620, 87930	87930	Three inference checkpoints and validation results
place_cube/lingbot-va/train90_base_job4139	300, 1500, 3000	3000	Three inference checkpoints and validation results
dish_utensil/pi0.5/train90_job4137	3557, 17789, 35579	35579	Three inference checkpoints and validation results
place_cube/pi0.5/train90_job4138	785, 3929, 7859	7859	Three inference checkpoints and validation results

Latest update: three checkpoints per completed run from September 15–16, 2026, spread across training and including the final checkpoint. See latest results for recorded metrics and evaluation limits.

The dataset repository is separate from this model repository. Original source licenses and terms continue to apply.