Kaiyang92/ced-fixed20-cache

CED-Small fixed-20 feature cache

This repository stores the precomputed CED-Small hidden-state cache used for the Follow-Mellow 42.5/43.5 reproduction. It contains derived tensors and indexing metadata only; it does not contain WAV, MP3, FLAC, or other raw audio files.

Fixed identity

• 110 shards (shard_000 through shard_109)
• 465,622 unique cached audio pairs; zero duplicate keys
• 184,578,827,186 bytes (171.902 GiB) across 660 cache files
• cache point: final_hidden
• tensor dtype: float16
• per-item maximum shape: 512 x 384
• fixed-20 policy: two on-disk 10-second segments are concatenated into one 20-second input before the CED forward pass
• CED model: mispeech/ced-small
• CED model revision: 06bb40c5ec089e96867ebc5246be02441f4a71e4
• source training JSON rows: 1,477,976
• source training JSON SHA-256: 7e56200bf0c1ae2a1f4e94715483e1d97c985e5eb58efad9a5ad22d0c1ebeb00
• identical row_to_global_cache.npy SHA-256 in every shard: beb4c93ea65e5722848260c4d0d5127e435c0ba2d96ad5aac805493e85f3d366

Absolute local paths preserved inside metadata.json are provenance strings; they do not need to exist on the download machine. Configure the cache root in the training runtime to point at the downloaded directory.

Layout

Each shard contains:

hidden.f16.dat
keys.jsonl
metadata.json
row_to_global_cache.npy
segment_lengths.npy
valid_tokens.npy

The complete cache was checked for shard count, metadata consistency, tensor shape and byte count, key count, duplicate keys, valid-token bounds, padding, and sampled finite values before upload.

Download

python -m pip install -U huggingface_hub
hf download Kaiyang92/ced-fixed20-cache \
  --repo-type dataset \
  --local-dir /path/to/ced_fixed20_cache

Do not rename or flatten the shard directories.