ChongkaiGao/planning_extra

DreamCan / svl_project — extra artifacts

Everything from the DreamCan (svl_project) research host that is not already in ChongkaiGao/DreamCan_Dataset or ChongkaiGao/DreamCan_Models: training checkpoints for every setting, eval results and media, paper/slide/video/website materials, project-made datasets, logs and the code snapshot.

Layout

Repo paths mirror the original absolute paths on the host with the leading / dropped, e.g. /mnt2/gck/checkpoints/wan_pusht_subgoal_lora/lora_step_0070000.safetensors → mnt2/gck/checkpoints/wan_pusht_subgoal_lora/lora_step_0070000.safetensors.

• Files ≥ 64 MB (checkpoints, datasets) are stored as-is.
• Smaller files are packed per directory into plain (uncompressed) tar bundles: <dir>/_bundles/tier<T>_<NNNN>.tar. Members are stored relative to <dir>, so tar -xf <dir>/_bundles/tier0_0000.tar -C <local>/<dir> restores the original tree.
• MANIFEST.jsonl.gz lists every uploaded item: repo path, kind (raw/tar), size, tier, and for tars the full list of original member paths.

Main roots:

prefix	contents
data/gck/svl_project/	code + configs + docs (working tree incl. uncommitted changes), eval/, outputs/, logs/, lower-policy / AR / IQL checkpoints under checkpoints/, robomimic data
mnt/gck/paper_materials/, mnt4/gck_planning/cover_v*/, dreamcan_slides_export/, dreamcan_figures_*/, mnt/gck/gck_website/	paper figures, video materials, cover videos, slides, website
mnt2/gck/checkpoints/, mnt2/gck/gck_wan_*/, mnt4/gck/gck_*ablation*/, mnt/gck/square_dp_aligned_30ep/	Wan2.2-TI2V-5B LoRA runs (LT line/a2r/cs/surround, LIBERO-Long/90, Push-T, Square data-size ablations: DP / evenly / dense / DP-aligned) with eval media + wandb
mnt/gck/svl_ckpt/, mnt2/gck/svl_ckpt/	LIBERO-Long lower-policy architecture/optimizer ablations (exp*)
mnt2/gck/lt_campaign_20260920/, mnt2/gck/lt_pilot_20261009/	Language-Table continuation campaign and block-integrity pilot
mnt/gck_planning/, mnt4/gck_planning/	real-robot tasks: box_folding (55/105 ep, Sirius r2), human_box_packing, YAM line — data (frames/*.h5, raw episodes), value fns, Wan LoRAs, policies
mnt4/gck/kitchen/, mnt4/gck/gck_square/, mnt2/gck/pusht/, mnt/gck/pusht_v2/	Franka Kitchen, Square, Push-T data / value fns / Robometer / policies
mnt/gck/vla_baselines/	openpi pi0.5 + LingBot-VA Push-T baselines
mnt/gck/svl_ops/	experiment ops: eval logs, reports, scripts
mnt4/gck_wandb_svl_project/	offline wandb runs
mnt/gck/hf_planning_extra/extras/	svl_project_git_bundle.bundle (full git history: git clone svl_project_git_bundle.bundle), uncommitted diff, project notes (_claude_memory/)

Tiers (upload priority)

0. docs / results / eval media / code / logs (small files)
1. final checkpoint of every training run
2. project-made datasets
3. low-priority data: converted LIBERO / LIBERO-90 hdf5, WGO (droid/homer/galaxea) data, old AR eval videos, archive
4. ~8 evenly spaced intermediate checkpoints per run (for picking best-eval steps)

Deliberately not uploaded (regenerable or public)

Precomputed RADIO/feature caches (wan_subgoal_radio_sequences, *radio_feats*, ar_subgoal_radio_sequences, iql_features, */radio/*.h5 next to frames/), episode_cache_* pickles, the Wan2.2 / Qwen / umt5 base models, public raw datasets (LIBERO raw, Language-Table raw), optimizer states, non-final intermediate checkpoints beyond tier 4, smoke/memtest runs, and all ManiLadder / pi05 material (separate project).