internvla-a15-v4b-joint-full

InternVLA-A1.5 — B1K joint v4b (withdrawn)

The weights that were here have been deleted (2026-09-09). Nothing in this repository should be evaluated, served, or fine-tuned from.

Held until then: step-005000 through step-025000, 5 checkpoints

Why

Every A1.5 round before internvla-a-series 63aa8af trained against a constant action chunk. Under an episode subset, LeRobotDataset.__getitem__ passed a subset-relative index to _get_query_indices, which clamps against absolute episode bounds; all 32 action deltas collapsed onto one index, so the target was a single frame tiled 32 times and carried no trajectory.

within-chunk max std (median)    subset 0.000000    no subset 0.351066

That makes every model-side number from this round void rather than merely superseded -- open-loop correlation and MSE, the serving prompt variants, and all rollout scores. Corpus measurements taken from the raw parquet never went through the loader and are unaffected.

The re-run on the fixed loader is at madokalif/internvla-a15-v5-joint.