otoha-project/gol-cache-v2

GOL feature cache v2 — audit and reconstruction guide

This audit covers gated repository otoha-project/gol-cache-v2 revision 2a370b14f94e16175d55d2afdd013e1257bae123. The repository has no card, schema, producer version, license, or reconstruction instructions.

Repository structure

The cache is one tar stream split into 18 raw parts:

• gol_cache.tar.part000 through part016: 21,474,836,480 bytes each;
• gol_cache.tar.part017: 7,535,349,760 bytes;
• total: 372,607,569,920 bytes.

All part boundaries and the total size are 512-byte aligned. Reconstruct in exact numeric order without decompressing the parts:

cat gol_cache.tar.part{000..017} > gol_cache.tar
tar -tf gol_cache.tar | head

Range inspection verified a valid tar header at byte zero, continuous member sequences across interior samples, and a complete final member followed by 13 zero blocks. The tail is therefore structurally terminated.

Sampled member schema

Members are grouped under eight-character speaker directories and stored as .pt files. Three prefix members were valid PyTorch ZIP serialization archives, little-endian, format version 3. Opcode-only inspection of the first data.pkl—without unpickling—found these keys:

speaker, utt_id, mel, phonemes, accent4ch, n_moras, moras, wav_path, text, duration, mos, transcription_source, and cache_version.

The inspected example has mel shape (1, 55, 80), phonemes shape (1, 1), accent4ch shape (1, 1, 4), transcription_source="gol_native", and cache_version=1. These are sample observations; a complete member manifest and tensor schema are still required.

Safe loading

PyTorch serialization contains a pickle metadata stream. Do not call torch.load on untrusted files with unrestricted unpickling. Validate provenance and hashes first; use map_location="cpu" and weights_only=True where the stored structure is compatible. See the official PyTorch torch.load security warning: https://docs.pytorch.org/docs/stable/generated/torch.load

Missing documentation

Before relying on the cache, record the exact source dataset revision, preprocessing code commit, PyTorch/torchaudio versions, mel parameters, phoneme and accent vocabularies, MOS model, expected tensor dtypes/shapes, and whether wav_path is relative to midralab/gol-dataset. No license was present; gating is not a substitute for explicit audio and derived-feature rights.