Modelpublic

ZhuoranChen/lingbot-va-mot-posttrain-libero-goal

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes

Model Card

lingbot-va-mot-posttrain-libero-goal

Full-parameter post-train of `lingbot-va-mot` on the LIBERO-Goal benchmark (10 tasks, original prompts), using real episode packing — multiple episodes concatenated into one joint forward pass per optimizer step, matching the lingbot-va paper's training regime. Supersedes `lingbot-va-mot-posttrain-libero-goal-gradaccum43`, which only matched the paper's token count per step (43 independent single-episode micro-forwards, gradients summed) rather than the paper's joint multi-episode attention.

Training config


base checkpoint	`checkpoints/lingbot-va-mot`
dataset	LIBERO-Goal (`data/libero_goal_lerobot`)
`freeze_backbone`	`False` — full-parameter fine-tune, all 10.0B params (30-layer video stream + action stream + embedders/heads)
steps	4000
learning rate	1e-5
world size	2 GPUs
packing	real episode packing (not grad-accum token-count matching) — each optimizer step processes several full episodes concatenated into ONE joint forward pass, with per-episode isolation enforced by a FlexAttention block mask (self-attention and text cross-attention)
`target_tokens_per_pack`	12,000
`gradient_accumulation_steps`	4 (now counts packs, not episodes) → ≈96K tokens/optimizer-step across 2 GPUs, matching the paper's ~100K
`window_size` (training)	randomly sampled per step, uniform in [4, 65)
`frame_chunk_size`	randomly sampled per step, uniform in [1, 5)
`guidance_scale` / `action_guidance_scale`	5.0 / 1.0
`num_inference_steps` / `action_num_inference_steps`	25 / 50
`snr_shift` / `action_snr_shift`	5.0 / 0.05
action representation	30-dim padded vector; native 7-dim LIBERO OSC action (xyz + euler + gripper) in slots 0–6, rest zero-padded
observation cameras	`agentview_rgb`, `eye_in_hand_rgb`

Architecture note: action stream width

This checkpoint (and its lingbot-va-mot base) uses a no-bottleneck MoT design: the action stream runs at the full video-stream width d_v = 3072 end-to-end (action_embedder: Linear(30 → 3072), action_proj_out: Linear(3072 → 30), and every per-block action module — action_attn1/2, action_ffn, action_norm2, action_scale_shift_table — is shape-identical to its video-stream counterpart). This differs from the lingbot-va paper, which describes the action stream operating through a narrower 768-dim bottleneck (30 → 768 → … → 768 → 30) rather than the full 3072-dim width used here. See `lingbot-va-mot`'s model card for how these action-stream weights were initialized before this fine-tune.

LIBERO-Goal closed-loop evaluation (checkpoint step 4000)

10 episodes/task, closed-loop rollout in the LIBERO-Goal env. attn_window=30 at eval (not 72 — training's window_size is always sampled below 65, so an eval-time window of 72 would be out-of-distribution relative to what the model ever saw during training):

Task	Success
open the middle drawer of the cabinet	10/10
put the bowl on the stove	10/10
put the wine bottle on top of the cabinet	9/10
open the top drawer and put the bowl inside	10/10
put the bowl on top of the cabinet	10/10
push the plate to the front of the stove	10/10
put the cream cheese in the bowl	9/10
turn on the stove	10/10
put the bowl on the plate	10/10
put the wine bottle on the rack	10/10
Total	98/100 (98.0%)

For comparison, the gradient-accumulation-only checkpoint scored 96/100 (96.0%) under the same protocol — real packing gives a small but clean improvement, with no regression on any task, and both are close to the paper's reported ~98.5%.

Repo contents

Only transformer/ (the fine-tuned weights) is included here. The original checkpoint directory also symlinks text_encoder/ (google/umt5-xxl), tokenizer/, and vae/ (Wan2.1 AutoencoderKLWan) from the shared base checkpoint — those are unchanged stock components and are not duplicated in this repo. Load them from `ZhuoranChen/lingbot-va-mot` or the public Wan2.1 release when using this checkpoint standalone.