JustANormalTinkerer/Japanese-Dialogue-Dataset

Japanese Dialogue Dataset

Deduplicated single-line Japanese dialogue mined from community subtitles.

I WILL UPLOAD A MORE CLEAN VERSION LATER.

Source (attribution — please keep)

• Corpus: Ajatt-Tools/kitsunekko-mirror, a mirror of subtitle files from kitsunekko.net/subtitles/japanese/, collected and contributed by volunteers, plus files mirrored from kitsunekko.
• This dataset republishes short dialogue lines only (no timing info, no audio/video). Subtitle texts belong to their respective authors/rights holders (fan subbers, broadcasters, streaming platforms). Provided for research / language-learning purposes.
• Closed-caption (-cc) subtitles were excluded: any file whose name contains [cc] / (cc) or a standalone cc token (e.g. .CC., -cc, _cc) was skipped, since CC tracks contain extra sound-effect / speaker annotations rather than clean dialogue.
• Categories covered: anime_movie, anime_tv, drama_movie, drama_tv, unsorted

How it was built

1. Walk subtitles/ for *.srt / *.ass / *.ssa (multi-encoding read: utf-8-sig → utf-8 → cp932 → shift_jis → euc-jp → latin-1).
2. Skip cc files (see above).
3. Parse properly per format:
   • SRT: block structure index / start --> end / 1..n text lines, timestamp-validated.
   • ASS/SSA: only Dialogue: events in [Events] (Format-aware split, Text = last field; Comment: ignored); override tags {...} removed; \N / \n → line break, \h → space.
4. One line per row: every cue split on newlines and - speaker dashes; HTML tags and leading dashes stripped; each row guaranteed to hold exactly one dialogue line (no embedded newlines).
5. Salvage pass (do not throw away dialogue) applied to every row:
   • Invisible / bidi formatting controls stripped (LRM, RLM, ZWJ, ZWNJ, BOM, bidi overrides).
   • Emoji removed from the line (📱, 😀, 🔥, ➡ …) — remaining text kept. Preserved on purpose: hearts ♡ ♥ ❤ 💕 💖 …, music ♩ ♪ ♫ ♬ ♭ ♮ ♯ 🎵 🎶 🎼, stars ★ ☆.
   • Leading speaker / stage tags stripped, repeatedly: （島田）お疲れ → お疲れ, （島田）（小声）… → ….
   • Stray / doubled brackets peeled: unbalanced ((text / text)) → text; balanced-but-doubled wrappers ((…)) / （（…）） → …. Single balanced wrappers like （笑） are kept as-is.
   • Rows that still look like a stage direction (bracketed name + decorative arrow) are dropped.
6. No language/content filtering by default: bilingual (ja-en) English lines and music-symbol-only lines (e.g. ♬～) are kept.
7. Deduplicated per split: exact-match dedup on the cleaned dialogue string within each split; first-seen provenance kept. (The same common phrase may legitimately appear in more than one split.)
8. Splits (one per category):
   • anime_tv: TV anime subtitles (subtitles/anime_tv/…).
   • anime_movie: anime film subtitles (subtitles/anime_movie/…).
   • drama_tv: TV drama subtitles (subtitles/drama_tv/…).
   • drama_movie: drama film subtitles (subtitles/drama_movie/…).
   • unsorted: everything else (subtitles/unsorted/…).
9. Columns:
   • dialogue: single cleaned Japanese dialogue line.
   • source: cleaned source title (parent show/film folder, whitespace-normalised), e.g. AKIRA.
   • category: anime_movie | anime_tv | drama_movie | drama_tv | unsorted.
   • source_file: original subtitle filename (provenance).

Use

from datasets import load_dataset
ds = load_dataset("JustANormalTinkerer/Japanese-Dialogue-Dataset")  # all 5 splits
print(ds["anime_tv"][0])
# or a single split:
anime = load_dataset("JustANormalTinkerer/Japanese-Dialogue-Dataset", split="anime_tv")

License / caveats

• No license is granted over the underlying subtitle texts; respect the original authors' rights.
• Stage directions like （笑い声） may occur; lines are deduplicated but not otherwise filtered.
• Might have a few english rows, filter as needed. I've kept them for personal use cases. (just drop anything that has -en in the source_file column)
• If you are a rights holder and want content removed, open an issue on the Hub repo.