Pref-VLA / SPT — Stage-A checkpoints (backup 2026-06-11)
Stage-A = preference-conditioned action training on the labeled source (taskA)
of the 0526 data (kaiwen2/robotwin-prefvla-0526),
QwenOFT backbone warm-started from StarVLA/Qwen3-VL-OFT-RoboTwin2-All
(steps_140000), 10k steps. Code/docs: GitHub Kaiwen-Hong/starVLA opd,
tag backup-0611 (r-preference/doc/0611-hf-backup-restore.md = restore runbook).
| run | framework | role |
|---|---|---|
pref_oftvqa_token_<cat>_10k |
QwenOFT_VQA (state_mode: token) |
main-method Stage-A: action + token-VQA cotrain. The frozen VQA branch is the Stage-B pseudo-labeler (predict_preference); also the warm-start for pref_stageb_main_* |
pref_oft_baseline_<cat>_10k |
QwenOFT (no VQA loss) |
ablation control + warm-start for pref_stageb_b0_* (Naive FT) |
Cats: height, orient, contact, place, hvlv. steps_10000 each, with
config.yaml / dataset_statistics.json / summary.jsonl sidecars.
NOTE for loading the token ckpts: build the model FROM THE YAML
(starvla_pref_stage_a_oftvqa_token_<cat>.yaml, so state_mode: token
initializes vqa_state_proj) then load_state_dict(strict=False) — see
r-preference/eval/gate_oftvqa.py.
Restore layout: results/Checkpoints/<run>/checkpoints/steps_10000_pytorch_model.pt,
sidecars at <run>/ root. Integrity: SHA256_MANIFEST.txt.