kaiwen2 commited on
Commit
e7a66cb
·
verified ·
1 Parent(s): 3a59abf

README: run naming + restore layout

Browse files
Files changed (1) hide show
  1. README.md +32 -0
README.md ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ ---
4
+
5
+ # Pref-VLA / SPT — Stage-B checkpoints (backup 2026-06-11)
6
+
7
+ Stage-B = pseudo-label-conditioned SFT on unlabeled taskB demos (QwenOFT, L1,
8
+ 14-D joint, chunk 50), warm-started from the Stage-A ckpts in
9
+ [`kaiwen2/prefvla-ckpts-stagea`](https://huggingface.co/kaiwen2/prefvla-ckpts-stagea).
10
+ Code/docs: GitHub `Kaiwen-Hong/starVLA` branch `opd`, tag `backup-0611`
11
+ (see `r-preference/doc/0611-hf-backup-restore.md` for the restore runbook).
12
+
13
+ ## Run naming
14
+
15
+ | prefix | meaning |
16
+ |---|---|
17
+ | `pref_stageb_main_<cat>` | SPT arm: token-Stage-A init + pseudo-label suffix (token-VQA labels) |
18
+ | `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) |
19
+ | `pref_stageb_b0_<cat>` | Naive-FT control: baseline-Stage-A init, NO preference suffix |
20
+ | `pref_stageb_yn_place` | ablation: token init + NO suffix ("VQA-only") |
21
+ | `pref_stageb_ny_place` | ablation: baseline init + vanilla-VLM labels (0.61 acc) |
22
+
23
+ `steps_1500` = the evaluated checkpoint everywhere; `steps_2500` additionally
24
+ for `main_contact_geom`, `main_hvlv_geom`, `b0_contact`, `b0_hvlv` (the
25
+ "more steps" experiment). Each run ships `config.yaml`,
26
+ `dataset_statistics.json` (denorm q01/q99 — required for serving), `summary.jsonl`.
27
+
28
+ Layout to restore: `results/Checkpoints/<run>/checkpoints/steps_N_pytorch_model.pt`
29
+ with sidecars at `<run>/` root. Serving: `deployment/model_server/server_policy.py`
30
+ + the `starvla_joint` bridge (identity reorder, `chunk_step=50`).
31
+
32
+ Integrity: `SHA256_MANIFEST.txt` (`OK <sha256> <bytes> <path>`).