CoolFace
Modelpublic

hz1919810/flashwam-arx-student-from-t3000

sourceHugging Faceotherupdated 24d agoView on Hugging Face
0likes
Model Card

Flash-WAM Distilled Student (ARX) — from teacher step 3000

Modality-aware consistency-distilled student of the LingBot-VA ARX world-action model, produced with Flash-WAM (official code implementation, paper-style 2000-step duration).

Teacher: LingBot-VA ARX SFT v2 split checkpoint step 3000 (160-episode train split). Distilled on the same 160 train episodes.

Contents

  • —target_student/transformer/ — EMA target (recommended for inference).
  • —online_student/transformer/ — trainable online student at step 2000.

Both are diffusers-format WanTransformer3DModel (config.json + bf16 safetensors, ~9.5 GB each).

Distillation recipe (per collaborator spec)

modeflashwam (video + action consistency + action-aware reg)
steps2000 (paper duration), save 500/1000/1500/2000
lr / betas / wd / warmup5e-6 / (0.9, 0.999) / 0 / 100
ema_decay0.995
loss / huber_chuber / 0.001
actionlossweight / actionawareweight1.0 / 0.01
numddimtimesteps (video / action)2 / 2
cfg range[2.0, 10.0]
effective global batch16 (4 GPU x accum 4)
final loss (video / action)0.2877 / 0.0120

ARX contract (unchanged from teacher)

3 cameras (camhigh, camleftwrist, camrightwrist); 256x256; K2 profile `video7p5hzaction15hzk2`; 14-D absolute joint target -> 30-D layout `[14,15,16,17,18,19,28,21,22,23,24,25,26,29]`; dataset q01/q99; envtype=none, actionperframe=8, attn_window=30.

Caveat

This is a fast (1-2 step/modality) student. Low distillation loss does not guarantee multi-step rollout fidelity or real-robot task success. Evaluate with rollout / real-robot before deployment; final selection (online vs EMA, and vs teacher) is the data owner's call.