Modelpublic

hmkang/wam_ctxpool_avg

sourceHugging Faceapache-2.0updated 10d agoView on Hugging Face

1likes1.4kdownloads

Model Card

WAM_DIT4DIT — context pooling (avg), RoboCasa kitchen, video-only

Wan2.2-TI2V-5B video DiT fine-tuned on RoboCasa (base recipe, effective batch 64: 4 GPU × per-device 8 × grad-accum 2), 4-latin history (nin=25 / nout=41, fdf=2), with average context pooling of the past latent frames inserted before block L.

folder	pooling	layer	steps
`L12/checkpoint-<step>`	avg	12	every 20k
`L3/checkpoint-<step>`	avg	3	every 20k

Weights (*.safetensors) + config.json + processor / experiment config only; optimizer state and training_args.bin are not included. Training code: https://github.com/HEMMO0208/wam (run_scripts/train/wam_dit4dit/compression/finetune_wam_dit4dit_robocasa_kitchen_ctxpool.sh).

Note: an earlier version of this repo held effective-batch-32 runs; those were removed. All checkpoints here are eff-64.