model_cache_code_step8000/DATA_ASSETS.md

Data Assets

There are three distinct data layers. Do not collapse them when uploading or restoring.

1. Original Raw Data

Original source data is PanTS:

/scratch/user/yuhwang/dataset/PanTS

Approximate size observed before packaging: 371G.

Structure:

• data/ImageTr: training images/volumes
• data/ImageTe: test images/volumes
• data/LabelTr: training labels
• data/LabelTe: test labels
• data/metadata.xlsx: raw metadata table
• repo: PanTS source README/license/download scripts

Raw data is packaged separately under pants_raw_train_test_20260523_102411.

2. VAE Latent Cache

Derived VAE cache used by the training dataloader:

/scratch/user/yuhwang/dataset/pants-captions-ldm/cache/wan22_pants_v2_softwin/latents

This stores pre-encoded Wan2.2 VAE latents as .safetensors. The train manifest references these via latent_path.

3. Text Cache

Derived text embedding cache:

/scratch/user/yuhwang/dataset/pants-captions-ldm/cache/wan22_pants_v2_softwin/text_embeddings

The main training manifest is:

captions_embedded/source_train.jsonl

Observed row count: 53,784.

Observed bucket counts:

• B-whole: 21,690
• B-CA: 11,365
• B-CAP: 10,980
• P-pan: 8,784
• B-abd-pelvis: 490
• B-abd: 475

The manifest fields include bucket_id, caption, latent_path, latent_shape, text_embedding_path, text_embedding_shape, and CT metadata fields.

Cache Archive Included Here

cache/wan22_pants_v2_softwin.tar.zst.part-* is a split tar+zstd archive of the full derived cache directory. It includes:

• latents/: VAE latent cache
• text_embeddings/: text embedding cache
• captions/: caption jsonl files
• captions_embedded/: embedded-caption manifests used by training
• manifests/: pre-embedding manifest layer
• cache_config.json