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.
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_jointbridge (identity reorder,chunk_step=50).
Integrity: SHA256_MANIFEST.txt (OK <sha256> <bytes> <path>).