Checkpoint Archive Manifest — kaiwen2/robomme_checkpoint_archive (HF, private)

Rescued from H200 (kevin@34.34.93.23) before access loss (2026-06-02). These were UNIQUE to H200 (not on lf). Uploaded directly H200→HF. Big mirrored dirs (data 1.5T, starVLA_runs, pref) were NOT archived (lf has its own copies / too large / not this project).

Repo layout: checkpoints/<EXP>/<step>/... (orbax ckpts) + data/collect_vr_r2/ + this manifest.

Download / restore

huggingface-cli download kaiwen2/robomme_checkpoint_archive --repo-type model \
  --local-dir ./archive --include "checkpoints/<EXP>/*"
# restore for serving/eval: place under runs/ckpts/mme_vla_suite/<EXP>/ ;
# ensure <EXP>/history_config.txt exists (archived if it was present) so serve_policy
# rebuilds the right architecture (see policy_config.create_trained_policy).

Checkpoints (priority order = upload order)

#	EXP (repo: checkpoints/)	size	steps	what it is / eval status
1	...prog-v2_method4_warmstart_subset_main	23G	19999, 39998	M4 prog-warmstart (S1 aux-only 20k → e2e warmstart). doc-v6: M4 eval was PENDING/UNCERTAIN → likely ONLY copy; needs 3-seed eval @39998. 5-task subset
2	...prog-v2_method4_s1_subset_main	12G	19999	M4 Stage-1 (aux-only, no action loss) — the S1 source for #1. 5-task subset
3	...hint-v2-prog-v2_hintfreeze15k_main	89G	10k–79999 (8)	hintfreeze (hint+prog, HINT_LAMBDA frozen after 15k). Full 16-task 80k. doc-v6 Job-B (≈12–14% @79999, conflicting). Paper baseline
4	...hint-v2-sem-v2-prog-v2_d1_refined_main	89G	10k–79999 (8)	D1 refined (3 aux heads + AUX_COND=1 aux→AdaLN, Xavier0.1, sem dropped from aux_feat). doc-v6: 79999 4-seed=14.0±2.3, 40000=12.5±2.0. Full 16-task 80k
5	...prog-v2-multibump_v11_sam2_subset_main	56G	5k–24999	v11 sem+prog combo + SAM2 dynamic tracking. 5-task subset (~27% @25k)
6	...prog-v2_V3_cyc5k_subset_main_main	34G	10k–24999	V3 CycFreeze (cycle-consistency, 5k cyc). 5-task subset
7	...hint-v2-sem-v2-prog-v2_V1_d2_seed7_subset_main_main_seed7	34G	10k,20k,24999	V1 D2 cross-attn, seed7. 5-task subset
8	v20_16task_phase2_bs384_main	45G	2k,4k,6k,8k	phase2 bs384 INTERMEDIATES (pre-kill). Final = ..._resume8k/1999 (already on lf). 16-task
9	cw_videorepick_dense	56G	2k–9999	MINE (doc-v8): channel-widen AUX_COND=2 dense ckpts (VideoRepick). Reproducible
10	cw_videorepick_ac2_c2	23G	10k, 19999	MINE: channel-widen VideoRepick (10k=26.7% / 20k=21% test pass@1). Reproducible

Data

• data/collect_vr_r2/ (29G): round2 filtered-BC collected trajectories (VideoRepick, policy self-rollouts on train eps, seeds 5-8, SUCCESS-filtered). npz per-step (demo prefix + act). (round1 = collect_vr already mirrored to lf as collect_vr_h2.)

NOT archived (and why)

• data (1.5T), starVLA_runs (1.1T, kevin's other VLA project), pref (241G): lf has its own copies (sizes differ slightly — if a SPECIFIC starVLA/pref artifact is unique+needed, flag it).
• smoke ckpts (v20_smoke_*), method3_subset (already on lf): disposable / mirrored.

Status: uploading (H200 /tmp/hf_upload.out). HF token was shared in chat → ROTATE after.