InternVLA-A1.5 — B1K joint v5 (full fine-tune)
This repository holds the v5 round trained with the fixed data loader
(internvla-a-series 63aa8af), started 2026-09-08.
What was removed, and why
An earlier round occupied this repository as step-005000 … step-040000.
Those checkpoints were deleted on 2026-09-09 because they were void, not merely
superseded.
Under an episode subset, LeRobotDataset.__getitem__ passed a subset-relative index
to _get_query_indices, which clamps against absolute episode bounds. Every one of
the 32 action deltas collapsed onto the same index, so the training target was a
single frame tiled 32 times and carried no trajectory:
within-chunk max std (median) subset 0.000000 no subset 0.351066
Nothing measured from those weights stands — open-loop correlation and MSE, the
serving prompt variants, and every rollout score. The same defect was present in the
earlier v3 and v4 rounds. It is fixed in 63aa8af, and this round is the re-run.
Corpus-side measurements taken from the raw parquet are unaffected; they never went through the loader.
This round
| base | A1.5-base — Qwen3.5-2B VLM + 460 M action expert |
| corpus | b1k/joint_v5, held-out split drawn by source demonstration |
| schedule | 109,582 steps, global batch 128, LR 5e-5 cosine → 5e-6, warmup 2,000 |
| action chunk | 32 generated, 16 executed |
| precision | bfloat16, gradient checkpointing on |
Step directories are uploaded as the round progresses.