README.md

navcache — M7 v7 derived caches

Caches produced by the M7 v7 pipeline from public NAVSIM / OpenScene data. Nothing here is raw sensor data and nothing here is model weights; every prefix is something this project computed and would otherwise have to recompute. Written 2026-08-21, when the cluster that produced it was retired.

prefix	what it is	why it is worth its bytes
feat_cache/dinov2/	206 GB of frozen DINOv2-base features over navtrain/navtest	GPU-hours of encoding
feat_cache/dinov3/	838 GB of frozen DINOv3 ViT-7B/16 features	the largest single artifact in the project
test_cache/	v1.1 navtest, v2.2 navtest, v2.2 navhard-two-stage metric caches	evaluator-pinned, needed to score anything offline
train_cache/navtrain_portable/	portable navtrain metric export, 103,288 shards — one per M7 index token	re-labelling a new trajectory vocabulary needs only this
train_cache/navsim_v2_2_navtrain_full/	the raw evaluator-pinned navtrain metric cache it came from	~76 min on 128 threads to rebuild, and only inside the pinned v2.2 environment
hydra_mdp_labels/	navtrain_k8192_seed0 offline PDM rule labels	~1,300 core-hours of candidate scoring — the most expensive derived artifact here
vocabulary/	the navtrain_k8192_seed0 trajectory vocabulary	the labels above are keyed to it and are meaningless without it
index/	navtrain / navtest / navhard parquet indices	defines the token set everything else is aligned to

Read the dinov3 manifest carefully. feat_cache/dinov3/<hash>/manifest.json is the storage-fields-repaired version. The pre-repair copy is retained as feat_cache/dinov3/validation/manifest.pre-storage-fields.json — do not use it to bind the cache.

Not here, on purpose. The underlying NAVSIM sensor blobs, logs and maps (~1.4 TB) are a public download and were not re-hosted. Model weights are not here either: DINOv3 is gated and stays with its publisher.

Checkpoints, run records and scores live in the private companion bucket AndyPJT/corewam_ckpts, whose ARTIFACT_INDEX.md documents both buckets in full.