File size: 1,589 Bytes
e7a66cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 | ---
license: mit
---
# Pref-VLA / SPT — Stage-B checkpoints (backup 2026-06-11)
Stage-B = pseudo-label-conditioned SFT on unlabeled taskB demos (QwenOFT, L1,
14-D joint, chunk 50), warm-started from the Stage-A ckpts in
[`kaiwen2/prefvla-ckpts-stagea`](https://huggingface.co/kaiwen2/prefvla-ckpts-stagea).
Code/docs: GitHub `Kaiwen-Hong/starVLA` branch `opd`, tag `backup-0611`
(see `r-preference/doc/0611-hf-backup-restore.md` for the restore runbook).
## Run naming
| prefix | meaning |
|---|---|
| `pref_stageb_main_<cat>` | SPT arm: token-Stage-A init + pseudo-label suffix (token-VQA labels) |
| `pref_stageb_main_<cat>_geom` | SPT arm with geometric/trajectory pseudo-labels (place/contact/hvlv 1.00/1.00/0.92; orient_geom 1.00) |
| `pref_stageb_b0_<cat>` | Naive-FT control: baseline-Stage-A init, NO preference suffix |
| `pref_stageb_yn_place` | ablation: token init + NO suffix ("VQA-only") |
| `pref_stageb_ny_place` | ablation: baseline init + vanilla-VLM labels (0.61 acc) |
`steps_1500` = the evaluated checkpoint everywhere; `steps_2500` additionally
for `main_contact_geom`, `main_hvlv_geom`, `b0_contact`, `b0_hvlv` (the
"more steps" experiment). Each run ships `config.yaml`,
`dataset_statistics.json` (denorm q01/q99 — required for serving), `summary.jsonl`.
Layout to restore: `results/Checkpoints/<run>/checkpoints/steps_N_pytorch_model.pt`
with sidecars at `<run>/` root. Serving: `deployment/model_server/server_policy.py`
+ the `starvla_joint` bridge (identity reorder, `chunk_step=50`).
Integrity: `SHA256_MANIFEST.txt` (`OK <sha256> <bytes> <path>`).
|