This card is based on a read-only audit of revision 4b43743fbe5f84589130587c30d625a8b0d95694. The repository is gated and contains visual-novel speech, but the source card does not identify the included titles, rights holders, extraction procedure, or license. Access approval is not a license; maintainers must document the lawful use and redistribution basis before downstream use.

Exact repository inventory

607 files totaling 7,019,273,829,197 bytes.
602 tar files totaling 7,015,165,429,760 bytes; all are 10,240-byte aligned and all 602 LFS payload hashes are unique.
597 logical tar game IDs. Five games are represented by _part1.tar / _part2.tar pairs; the sampled pair consists of two independently terminated archives.
metadata.tsv: 1,676,537,850 bytes and 7,405,094 rows.
metadata.json: 2,431,857,412 bytes and 7,405,094 rows.

The JSON and TSV were compared in full. After normalizing the TSV parser's 1,055 null empty fields to the JSON representation "", their ordered rows and both multiset directions have zero differences across all five columns. Neither contains the omitted game described below.

The previous YAML value dataset_size: 105000000000 is not the repository payload size and should be removed or relabeled. The exact LFS-backed repository is about 7.019 TB (decimal).

Global metadata audit

The TSV strictly parses to the statistics stated in the original card: 596 games, 19,349 speakers, 10,654.322 hours, median 4.552 s, and a 0.3–58.6987 s range. Every path has exactly game_id/speaker/file.wav, and its first two components match the row.

Quality and consistency findings:

1,055 missing transcripts;
156,606 rows without detected Japanese script;
121 replacement-character rows;
21,848 suspected lost-newline strings matching Japanese-character + n + Japanese-character;
112 @ruby markup rows and 5,791 outer-whitespace rows;
9,060 repeated audio-path groups with 10,877 rows beyond the first;
2,400 same-audio groups (5,152 rows) with competing transcripts;
218,226 repeated non-null transcript groups with 1,081,415 rows beyond the first.

The source data is unchanged. The audit preserves all competing rows in its conflict table. The conservative source manifest excludes missing text and entire ambiguous audio-path groups, and collapses only repeated paths whose transcript agrees. It contains 7,390,766 rows; 7,206,961 pass the stricter text checks.

Recovered omitted game

One complete archive, FD8F9FE19B9D9394B88F5866EE17DB03.tar, has no corresponding game ID in either global metadata file. Its live index.json contains only {}, but index.bak.json contains 22,225 source rows.

The complete 24,625,346,560-byte tar was read and hashed:

all 22,225 index rows resolve one-to-one to 22,225 WAV members;
zero missing, unreferenced, unsafe-path, invalid-header, or duplicate-SHA audio files;
59 speakers, of which 33 are new relative to the global TSV;
38.706225 hours; 0.5671–26.2954 s, median 6.0954 s;
all files are 44.1 kHz, mono, PCM format 1, 32 bits per sample;
15,693 paths resolve through the speaker-hash directory and 6,532 through the original source directory; both mappings are recorded rather than normalized away;
21,910 rows are ready and 315 punctuation/non-Japanese rows remain for review.

Adding the recovered overlay to the untouched global metadata yields 7,427,319 rows, 19,382 distinct speakers, and 10,693.028245 hours. The conservative combined Parquet has 7,412,991 rows; its high-confidence view has 7,228,871 rows. Each recovered row carries an actual audio-content SHA-256, while global TSV rows are explicitly marked audio_validation="metadata_only".

The audit keeps the complete conservative set without duplicating almost the entire table in a second file. Selecting quality_status = 'ready' produces the 7,228,871-row high-confidence view.

Tar observations

Bounded range checks of a normal archive and both halves of a split game found valid tar starts and standard zero-block endings. A normal prefix sample and every recovered orphan WAV used 44.1 kHz mono PCM32. This is not a substitute for a format scan of all 7 TB.

The tar layout uses a global index.json plus nested WAV files rather than conventional WebDataset sample-key groups. The original library:webdataset declaration and the declared audio/metadata features should therefore be verified against the actual loader. A future Parquet publication should provide a directly viewable metadata layer.

Machine-readable audit summaries are available under audit/.

Required provenance and rights work

Document, at minimum, the included title list, source acquisition method, speaker/game hash derivation, transcript extraction and markup normalization, audio conversion settings, exclusion rules, and rights for redistribution and model training. Until then, do not infer authorization from the repository gate or from the existence of a download.

Downloads last month: 51

Total file size:

7.02 TB