Superseded — intermediate checkpoint (step 7 500 of 30 000)
This repository holds an in-progress checkpoint that was uploaded while training was
still running: step 7 500 of a 30 000-micro-step run of π₀.₅ with recurrent memory
(64 memory tokens, TBPTT K=2) on
tttonyalpha/so101_pick_cube_300_task3.
The run has since finished. Use the final weights instead:
| Checkpoint | TBPTT K |
|---|---|
avanturist/pi05-mem-so101-pick-cube-color-m64-k2-tbptt | 2 — the finished version of this run |
avanturist/pi05-mem-so101-pick-cube-color-m64-k8-tbptt | 8 |
Those repositories carry full model cards: task, training configuration, file inventory, and the caveat that normalization statistics live in the dataset rather than in the weights.
The files here are kept so that existing links do not break. step-007500/ is a quarter of
the way into the schedule, at a point where the cosine learning rate had not yet decayed;
it is not a checkpoint anyone should be evaluating.
Code: avanturist322/muVLA-pi0.5.