Hugging Face's logo

Datasets:
ayousanz
/
c4-ja-text-audit

Tasks:
Text Generation
Modalities:
Tabular
Text
Formats:
parquet
Languages:
Japanese
Size:
10K - 100K
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

C4 Japanese text-only export — audit

The audited source, ayousanz/c4-ja-text, is a text-only derivative of the Japanese portion of allenai/c4. It contains 1,024 Git LFS shards totaling 824,615,119,625 bytes. All declared shard ordinals 00000..01023 are present, every LFS SHA-256 is unique, and no duplicate or missing shard ordinal was found at revision f20a06fede96a076fedd59cb67dbe13f87c4d675.

Important format limitation

Despite the .json.txt suffix, sampled payloads are plain text, not JSON or JSONL. The export preserves newlines from source documents but provides no unambiguous document delimiter, document ID, URL, timestamp, or source metadata. Treating each physical line as one example would split multi-line web documents and should not be presented as a faithful C4 record reconstruction.

For a Viewer-ready derivative, regenerate from the upstream C4 rows and write Parquet with at least text, stable source/example ID, URL, timestamp, source split/shard, and a content hash. Do not infer document boundaries from this text export.

Bounded quality sample

A deterministic byte-range audit sampled two 2 MiB windows from 16 evenly spaced shards: 67,108,864 bytes, or about 0.0081% of the payload.

  • 29,389,961 decoded characters; 58.89% were Japanese-script characters.
  • 2,889 U+FFFD/replacement characters were observed (0.0098% of decoded characters; range-edge decoding contributes only a small bounded portion).
  • 77,942 normalized long-line occurrences produced 77,940 unique hashes.
  • Two exact long-line duplicate occurrences beyond the first were observed.
  • The sample had 268,627 line breaks and no blank-line boundary, reinforcing that the newline stream is not a safe document delimiter.

These are sample statistics, not full-corpus counts. Do not declare a row-count size category until a lossless regeneration, because this export cannot recover its own row count.

Machine-readable inventory and samples are under manifests/ and analysis/. This audit derivative has not modified the source repository and does not redistribute the 824.6 GB source payload.
Downloads last month
13
Total file size:
2.98 MB