HY memory-compression: fresh shared-rollout training (scale16-shared160-fresh-v1).
The compressor is initialized from the original pretrained HY assets (copied first 18 blocks) and fresh
queries/LayerNorm/residual MLP; no trained compressor seeds this run. Objective: one on-policy trajectory
per example supervising every 20P window (rollout warmup 40P -> 80P -> 160P), batch 16 on 16 GPUs,
cosine LR with a fixed endpoint of 800 updates. Code: branch impl/hyworld-shared-fresh-v1 of
https://github.com/Nicholas0228/video-compress (docs/HYWORLD_SHARED_ROLLOUT.md).
This dedicated private repository keeps one verified latest resumable checkpoint (latest/, LATEST.json),
up to four pending evaluation snapshots (eval/, every 50 updates), permanent weights in milestones/
(untrained step 0, every 500 updates, final) and the best evaluated weights. Old checkpoint history is
squashed after publication. Training data: dataset Nicholas0228/hyworld-memory-long65-prefixes@19fc8686;
evaluation data: Nicholas0228/hyworld-memory-scale16-prefixes@1e07e1eb.