Rootport/nz-traditional-haiku

Traditional Japanese Haiku Dataset (Edo–Meiji Era)

⚠️ Work in Progress — Pre-Release Draft

This dataset is not yet ready for general use. It is being uploaded primarily as a personal backup snapshot during active development. Schema, annotations, and documentation may change without notice. Approximately 22% of records (~2,800) are still flagged annotation_status = "needs_review" and have not yet undergone human review.

If you arrived here unexpectedly, please check back later — this notice will be removed when the dataset is ready for downstream use.

このデータセットは まだ製作中 です。現時点では作者本人のバックアップを兼ねた公開であり、スキーマ・アノテーション・ドキュメントは予告なく変更される可能性があります。約 22%(2,800 件)のレコードに annotation_status = "needs_review" が付いており、人間レビュー未反映 の状態です。

---

Dataset Summary

A machine-readable corpus of 12,787 classical Japanese haiku from seven public-domain poets spanning the late Edo to early Showa eras (1644–1959). Each record carries:

• Original source-text spelling (text) preserving historical orthography
• Modern hiragana reading (kana) and 5-7-5 segmentation (segments, mora_counts)
• Seasonal classification (season) and dual-form kigo annotation keeping both the source-text spelling AND the standard saijiki headword
• Form metadata (form, is_regular_575) distinguishing fixed-form haiku from free verse (Santōka, Hōsai)
• Author metadata with birth/death years and per-poet style tags

Designed primarily as training data for haiku generation models — including diffusion-based approaches — where the prompt input may vary from formal saijiki headwords (時雨 / しぐれ) to descriptive natural language (寒い夜).

概要

江戸後期から昭和初期(1644–1959)までの著作権切れ俳人 7 名・計 12,787 句 をアノテーションした機械可読データセット。各レコードに現代仮名読み・モーラ分割・季語(本文表記 + 歳時記見出し語の二重保持)・作者メタを付与。俳句生成モデル(Diffusion 系含む) の訓練データを主用途として設計。

---

Dataset Structure

Data Fields

Field	Type	Description
id	string	Unique ID, format author_slug_NNNNNN (e.g. matsuo_basho_000001)
text	string	Original haiku text, preserving source orthography (旧仮名遣い含む)
text_spaced	string	Text with half-width space at 5-7-5 boundaries
kana	string	Modern hiragana reading, segments joined by half-width spaces
normalized_kana	string	Normalized hiragana for training (typically identical to kana)
segments	list[string]	Reading split into 5-7-5 segments (1–3 elements for free verse)
mora_counts	list[int]	Mora count per segment; sums to total_mora_count
total_mora_count	int	Sum of mora_counts
form	string	One of 5-7-5, 5-7-5-extra, free, unknown
is_regular_575	bool	True iff mora_counts == [5, 7, 5]
season	string	One of spring, summer, autumn, winter, new_year, nonseasonal, unknown
kigo	list[dict]	Seasonal words; each entry has word, kana, canonical_word, canonical_kana, season, category
author	string	Poet name in kanji
author_kana	string	Poet name in hiragana
author_birth_year	int
author_death_year	int
source	string	Source publication
mood, mood_ja	list[string]	Mood labels (currently placeholder ["unknown"] / ["不明"])
scene, scene_ja	list[string]	Scene/motif labels; scene_ja populated, scene placeholder
style, style_ja	list[string]	Per-author default style tags
rights	dict	{text_status: "public_domain", annotation_license: "cc-by-4.0"}
annotation_status	string	One of ai_generated, needs_review, verified

Mora Counting

• Diphthong glides (ょ・ゅ・ゃ・ぁ etc.) count as 0 mora.
• Geminate (っ), syllabic n (ん), and long mark (ー) each count as 1 mora.
• Example: 「ちょうちん」 = ち + (ょ) + う + ち + ん = 4 mora.

Example Instance

{
  "id": "matsuo_basho_001234",
  "text": "古池や蛙飛び込む水の音",
  "text_spaced": "古池や 蛙飛び込む 水の音",
  "kana": "ふるいけや かわずとびこむ みずのおと",
  "normalized_kana": "ふるいけや かわずとびこむ みずのおと",
  "segments": ["ふるいけや", "かわずとびこむ", "みずのおと"],
  "mora_counts": [5, 7, 5],
  "total_mora_count": 17,
  "form": "5-7-5",
  "is_regular_575": true,
  "season": "spring",
  "kigo": [
    {"word": "蛙", "kana": "かわず",
     "canonical_word": "蛙", "canonical_kana": "かわず",
     "season": "spring", "category": "animal"}
  ],
  "author": "松尾芭蕉",
  "author_kana": "まつおばしょう",
  "author_birth_year": 1644,
  "author_death_year": 1694,
  "rights": {"text_status": "public_domain", "annotation_license": "cc-by-4.0"},
  "annotation_status": "ai_generated"
}

(Illustrative; the actual record corresponding to "古池や" has a different ID.)

Loading

from datasets import load_dataset

ds = load_dataset("Rootport/nz-traditional-haiku")
print(ds)

Or, with raw file access:

import json
with open("haiku_master.jsonl", "r", encoding="utf-8") as f:
    records = [json.loads(line) for line in f]

---

Authors

Slug	Name (kanji)	Name (kana)	Birth–Death	Count	Style
matsuo_basho	松尾芭蕉	まつおばしょう	1644–1694	4,231	classical / lyrical
yosa_buson	与謝蕪村	よさぶそん	1716–1784	867	classical / ornate / lyrical
kobayashi_issa	小林一茶	こばやしいっさ	1763–1828	1,098	classical / colloquial / humorous
masaoka_shiki	正岡子規	まさおかしき	1867–1902	3,745	classical / observational / plain
takahama_kyoshi	高浜虚子	たかはまきょし	1874–1959	1,736	classical / observational
taneda_santoka	種田山頭火	たねださんとうか	1882–1940	700	free verse / colloquial / lyrical
ozaki_hosai	尾崎放哉	おざきほうさい	1885–1926	410	free verse / colloquial / philosophical

Sources

Source texts are compiled from public-domain editions:

• Matsuo Bashō: 芭蕉俳句全集 (Yamanashi Prefectural University)
• Yosa Buson: ネット歳時記「きごさい」
• Kobayashi Issa: JANIS 一茶発句全集 + user-curated popular haiku
• Masaoka Shiki: 青空文庫 (Aozora Bunko)
• Takahama Kyoshi: 青空文庫
• Taneda Santōka: 青空文庫・草木塔
• Ozaki Hōsai: 青空文庫・尾崎放哉選句集

All source texts are out of copyright in Japan (latest death year 1959 + 70-year term).

---

License and Rights

• Original haiku text (the text field, in source orthography): Public Domain. All seven poets died before 1960; Japanese copyright has lapsed.
• Annotations (readings, segments, mora counts, kigo classification, seasonal tags, scene labels, and all other added metadata fields): Creative Commons Attribution 4.0 International (CC BY 4.0).

The schema's rights field on each record encodes this dual structure explicitly.

Annotation Methodology

• Readings (kana, segments) were produced by direct LLM annotation (Claude), not by general-purpose morphological analyzers such as pykakasi. An early experiment using pykakasi on a 410-record subset produced numerous systematic errors (e.g. 人 → にん, 月 → がつ, 哉 → や) that would have cost more in manual correction than the additional data was worth.
• For fixed-form poets (Bashō, Buson, Issa, Shiki, Kyoshi, early Santōka, early Hōsai), readings that fit the 5-7-5 pattern are preferred when ambiguous (e.g. 庵 = いお rather than いおり when the alternative would overflow the segment).
• For free-verse poets (later Santōka, later Hōsai), natural readings are preferred without forcing the 5-7-5 fit.
• Seasonal words (kigo) carry both the source-text spelling AND the standard saijiki headword form (canonical_word / canonical_kana). This dual form lets downstream models handle both descriptive prompts (「寒い夜」) and formal saijiki prompts (「時雨」) robustly.

---

Limitations and Known Issues

This is a preliminary release. Specific caveats:

1. 22% of records (2,800) carry annotation_status = "needs_review" and have not yet been human-reviewed. Many are likely deletion candidates: section headers, biographical metadata, editorial notes, quoted poems by other poets, and similar non-haiku content that survived from the source publications. A reviewer-input pass is planned before the v1.0 release.
2. Bashō hiragana-reading rows: roughly 1,180 records (mostly with IDs ≥ matsuo_basho_002500) are pure-hiragana phonetic versions of adjacent kanji-form haiku, retained from the source publication. They are intentionally preserved as separate records because hiragana-only training data is useful for downstream generation models. Be aware that pairs of records may share semantic content.
3. Source quality varies by author. The Shiki collection (青空文庫) includes substantial non-haiku material from the poet's late-period literary criticism — quoted classical poems by other authors, prose fragments, table-of-contents headers. NR rates in those sections exceed 95% and most entries are deletion candidates.
4. Mood labels are placeholders. mood = ["unknown"] / mood_ja = ["不明"] for every record. A separate downstream LLM annotation pass is planned.
5. Scene labels are partial. scene_ja (Japanese) is populated; scene (English) is placeholder pending downstream translation.

---

Citation

@dataset{nz_traditional_haiku_2026,
  title  = {Traditional Japanese Haiku Dataset (Edo--Meiji Era)},
  author = {Rootport},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/Rootport/nz-traditional-haiku},
  note   = {Pre-release draft. 12,787 records from 7 public-domain poets.}
}