run_wan14b_held.sh: correct stage-1/D to WANTRACK_FREEZE_HEAD=0 (train the
track head through stage-1, matching the 1.3B run D which never froze it and
whose head plateaus only by ~step 4700/4800). Was 1.
Overfit + stage-1 configs (14B):
- synth_sparse_random_14b_{lr2e5,bs8,bs16}: random-warmstart overfit variants;
bs16 (global bs 16, lr 1e-4) is the stable one whose ckpt-1200 merge scored
best on gradient-SNR (track_encoder.proj 0.530 / snr 0.969).
- openvid_sparse_14b_merged_fixed2000 + stage1_16node: stage-1 from the merged
init (global bs 128, lr 1e-5, 4800 steps); 16-node variant = 64 GPU, accum 2.
gradient_flow_analysis.py: register the 14B fun-control and bs16-1200 merged
inits for SNR comparison.
720p pipeline: preprocessing_datasets.py drops the persistent decoded-pixel
reference (1.34GB/clip at 720p) after each item so preprocess RSS stays bounded;
run_preprocess_720p_held.sh + verify_{720p_parquet,row_integrity}.py +
check_720p_latents.py to re-encode OpenVid at 1280x720 and verify index/track
alignment against the 480p set; smoke_720p configs; acquire_best.sh node grabber.
shaoxiongduan
36d6f8785e