haolunwang/baselines

/group/ycyang/whl baseline archive

Recorded 2026-11-09 before deleting local copies to free /group. Repo: haolunwang/baselines (this repository is public).

Two groups:

1. Official weights. Already published upstream. This file records the source repo. The multi-GB weight files were not copied here, because several include Llama-2 weights and the originals can be downloaded again. Local directories were deleted after this note.
2. Local ALOHA comparison runs (UniVLA / MVP-LAM / OpenCLAP). These are not on a public hub. Their checkpoints are uploaded under the same folder names, then deleted locally.

Left on disk on purpose: univla_multiview (own LIBERO work), recap_data / recap_experiment / robotwin_recap (active RECAP, Oct–Nov 2026), gemma-3-270m and siglip2-so400m-patch14-224 (downloaded 2026-10-28), raw/lerobot datasets (drawer, put_red_cube_into_blue_cup, aloha_multiview, libero_*_lerobot, rlds), and .cache/huggingface/datasets (~233G parquet cache).

Common ALOHA fine-tune recipe unless a row says otherwise: global batch 64 (per-device 8, grad accum 2), LoRA r32 dropout 0, lr 3.5e-4 (LoRA lr 5e-5, head lr 3e-4), window 12, downsample stride 3, target anchored_delta_qpos (right arm qpos 7:14), binary gripper, image aug, single camera, proprio-concat action decoder. Benchmark is the real ALOHA task named in dataset_name.

Official baselines (re-download, weights not re-hosted)

Local path	What	Params / benchmark	Upstream
openvla-7b (15G)	OpenVLA-7B base	Prismatic + Llama-2-7B, OXE pretrain	openvla/openvla-7b
openvla-oft/ (60G, 15G × 4)	OpenVLA-OFT official LIBERO fine-tunes	suites libero_10, libero_goal, libero_object, libero_spatial	moojink/openvla-7b-oft-finetuned-libero-{10,goal,object,spatial}
mvp-lam (932M)	MVP-LAM latent-action model	cross-view LAM, OXE + EgoExo4D	JM-Lee/mvp-lam
mvp-lam-7b-bridge-pt (15G)	MVP-LAM VLA Bridge pretrain	prism-dinosiglip 7B + MVP-LAM	JM-Lee/mvp-lam-7b-bridge-pt
mvp_lam/{libero_10,libero_goal,libero_object,libero_spatial} (57G)	MVP-LAM LIBERO fine-tunes, used for local LIBERO-PRO	LoRA r32, lr 2.5e-4, batch 8 × accum 2, 30k steps, codebook 16, image aug. libero_10 decoder action_decoder-30000.pt	JM-Lee/mvp-lam-7b-224-libero
univla-7b (15G)	Official UniVLA-7B	task-centric latent actions	qwbu/univla-7b
univla-latent-action-model (5.6G)	Official UniVLA LAM	lam-stage-1.ckpt, lam-stage-2.ckpt	qwbu/univla-latent-action-model
nora-1.5-libero (7.5G)	NORA-1.5 LIBERO	action chunk 5, VLM declare-lab/nora-long, FAST tokenizer	declare-lab/nora-1.5-libero
vla-jepa/ (16G)	VLA-JEPA LIBERO checkpoint plus backbones	LIBERO/checkpoints/VLA-JEPA-LIBERO.pt (5.8G); Qwen3-VL-2B-Instruct (4G); vjepa2-vitl-fpc64-256 (6G)	ginwind/VLA-JEPA (LIBERO subfolder), Qwen/Qwen3-VL-2B-Instruct, facebook/vjepa2-vitl-fpc64-256

notes/openvla-7b-local-code/ keeps the 2026-09-05 local modeling_prismatic.py / config.json and the *.back.20260905_232808 originals. The 15G safetensors were the public checkpoint and were not copied.

nora-1.5/ on disk was only the git clone (~503K) and was left in place.

Local ALOHA runs uploaded under the same folder name

OpenCLAP / QwenPIKM, primary view only (use_wrist_image: false), backbone ./pretrained/clap-s3-l32 (Qwen3-VL-4B, 36 layers, hidden 2560), DiT 16 layers, seed 42, AdamW, batch 8 × grad accum 2, image aug. Checkpoints are full pytorch_model.pt (~12G each). final_model/pytorch_model.pt is the last step.

Folder	Task	Train setup	Checkpoints
aloha_drawer_clap_s3_h8_qwen3vl4b_l16_sv	ALOHA drawer, data_mix=aloha_drawer	action horizon 8, 2500 steps, save every 250, lr base 2.5e-5 / VLM 3e-5 / action 3e-5	steps 250…2500
aloha_drawer_openclap_ckpt	same drawer run as the l32 folder below; only the step dumps	horizon 32, 5000 steps, save every 500	steps 500…5000
aloha_drawer_clap_s3_l32_qwen3vl4b_l16_sv	ALOHA drawer	horizon 32, 5000 steps, lr base 2.5e-5 / VLM 1e-5 / action 1e-4	final_model only (last step); intermediates are in aloha_drawer_openclap_ckpt
aloha_put_red_cube_clap_s3_l32_qwen3vl4b_l16_sv	ALOHA put red cube into blue cup	horizon 32, 5000 steps, same lrs as drawer l32	steps 500…5000 + final_model

UniVLA / MVP-LAM ALOHA fine-tunes use the common recipe above.

Folder	Init	LAM	Task	Steps	View
aloha_drawer_finetune/...drawer_v1_binary	own multiview VLA hf_step-025000	own multiview LAM step 80000	drawer	4000, save 200	VLA single, LAM multiview
aloha_drawer_finetune/...drawer_v1_binary_6k	same	same	drawer	6000	VLA single, LAM multiview
aloha_drawer_finetune_univla/...drawer_v1_binary	univla-7b	official lam-stage-2.ckpt	drawer	4000	both single
aloha_drawer_finetune_mvplam/...drawer_mvplam_v1_binary	mvp-lam-7b-bridge-pt	mvp-lam	drawer	4000	both single
aloha_hang_tape_finetune/...hang_tape_v1_3200	own multiview VLA hf_step-025000	own multiview LAM step 80000	hang tape	3200	VLA single, LAM multiview. Extra step-* dirs next to the run are adapter/optimizer dumps
aloha_hang_tape_finetune_univla/...univla_sv	univla-7b	official LAM stage-2	hang tape	3200	both single
hang_tape_univla_ckpts	exported steps of the univla_sv hang-tape run		hang tape	step-000200 … step-002800	adapter + optimizer, not a full 7B merge
aloha_hang_tape_finetune_mvplam/...mvplam	mvp-lam-7b-bridge-pt	mvp-lam	hang tape	3200	both single
hang_tape_mvplam_ckpt	exported steps of the MVP-LAM hang-tape run		hang tape	step-000200 … step-002800	adapter + optimizer
aloha_pick_cube_finetune_mvplam/...pickcube_mvplam_4gpu	mvp-lam-7b-bridge-pt	mvp-lam	put red cube into blue cup	4000	both single, 4 GPU
aloha_pick_cube_finetune_mvplam_ckpt	exported steps of that pick-cube run		put red cube	step-000200 … step-003800	adapter + optimizer