Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
carrier: large_string
scope: large_string
fields: large_string
notes: large_string
-- schema metadata --
pandas: '{"index_columns": [], "column_indexes": [], "columns": [{"name":' + 523
to
{'country': Value('large_string'), 'fields': Value('large_string'), 'postcode_format': Value('large_string'), 'notes': Value('large_string'), 'source': Value('large_string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/parquet/parquet.py", line 220, in _generate_tables
                  yield Key(file_idx, batch_idx), self._cast_table(pa_table)
                                                  ~~~~~~~~~~~~~~~~^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/parquet/parquet.py", line 156, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              carrier: large_string
              scope: large_string
              fields: large_string
              notes: large_string
              -- schema metadata --
              pandas: '{"index_columns": [], "column_indexes": [], "columns": [{"name":' + 523
              to
              {'country': Value('large_string'), 'fields': Value('large_string'), 'postcode_format': Value('large_string'), 'notes': Value('large_string'), 'source': Value('large_string')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

World Place Names — veygrit/world-place-names

世界の行政階層地名 + 郵便番号 + 住所翻訳/ローマ字化/字形変換/曖昧性解消/住所フォーム/入力補助/ジオコーディングの事前学習データ。 住所入力補助・正規化・翻訳パイプライン用。毎日差分更新、履歴保持。

翻訳は国際英語ピボット設計:各国母国語表記 ↔ 国際英語(米国の国内書式はUSPS Pub.28略称で別途)。 漢字圏は同名異地・同字異読み・繁体↔簡体↔日本新字体の3層で対応。

現在の規模(スナップショット 2026-09-23、実測値)

地名・郵便マスター

ファイル 行数 内容
data/places.parquet 287,617 country 252 / admin1 3,865 / admin2 47,643 / city 235,857
data/names.parquet 2,224,214 多言語別名(preferred / short / colloquial / historic / valid_from / valid_to)
data/wikidata_names.parquet 2,187 Wikidata 公式名称(P1448、開始/終了日付き)
data/postal/geonames_postal.parquet 1,826,904 世界郵便番号 121カ国(日次更新ソース)
data/postal/jp_postal.parquet 124,522 日本:郵便番号 + 漢字 + カナ + ヘボン式ローマ字(97.9%付与)

学習データ(data/training/)

ファイル 行数/件数 内容
place_name_pairs.parquet 873,046 / 563言語 母国語 ↔ 国際英語 地名ペア(ru 79k / zh 54k / fa 39k / uk 37k / ja 30k …)
ja_address_rome_pairs.jsonl 119,999 日本郵便ヘボン式:郵便番号 + 漢字住所 ↔ ローマ字(単語境界保持)
sft_address_translation.jsonl 450,799 messages形式SFT(jp_kanji_to_rome 120k / place_native_to_en 250k / jp_rome_to_kanji 30k / place_en_to_native 50k / address_carrier_format 800)
usps_abbreviations.csv 142 USPS Pub.28:街路61 / ユニット12 / 方位8 / 州61
carrier_constraints.csv 5 1行文字数・行数・大文字化の初期値(要検証のガイド値)
han_script_pairs.parquet 950 字形変換ペア:簡体↔繁体 840 + 日本新字体↔繁体 110(同一場所の実表記)
han_character_variants.parquet 14,708 Unihan kSimplifiedVariant/kTraditionalVariant(字単位:国→國, 万→萬, 学→學)
homograph_hanzi_places.parquet 7,388 同字異地:同一漢字表記が別の場所・別の英語名を持つ
han_reading_variants.parquet 25,610 同字異読み:同一漢字町名が複数読みを持つ 4,102字(例:一宮=イチノミヤ[千葉]/イチミヤ[静岡])
sft_han_tasks.jsonl 29,288 han_script_convert 1,900 / homograph_disambiguate 7,388 / ja_kanji_reading 20,000
homograph_crosscountry.parquet 70,928 同名異地(全言語):15,112文字列が複数国に跨る(例:Springfield=CA/JM/AU/GB)。疑似言語タグ(link/post/unlc/wkdt)はパイプラインで除外済み(旧88,994から18,066行除外)
typo_correction_pairs.parquet 94,711 スペルミス訂正ペア(edit-distance-1 + OCR/キーボード混同)。疑似言語タグ(link/post/unlc/wkdt = URL・郵便参照コード・wikidata id)はパイプラインで除外済み(369,692行の代替名を除外して生成)
languages_scripts_by_country.parquet 29,127 国別言語・文字採用表(248カ国、ISO 15924 + ローマ字化可否)。疑似言語タグ行は除外済み
sft_disambig_typo.jsonl 180,629 crosscountry_disambiguate 70,928 / crosscountry_candidates 14,990 / typo_correct 94,711
address_forms_by_country.parquet 48 国別フォーム欄順序 + 郵便番号正規表現(JP=postcode→prefecture→city→ward→sublocality→building→room)
carrier_forms.parquet 5 配送業者別の欄構成(USPS/FedEx/UPS/DHL/JapanPost-EMS)
ec_forms.parquet 9 主要ECプラットフォームの住所フォーム欄(Amazon各ドメイン/Rakuten/eBay/Taobao/Coupang)
sft_address_forms.jsonl 30,062 address_to_form 30,014(JP実かな+US実郵便データ)/ form_schema_query 48
sft_postcode_assist.jsonl 40,000 postcode_assist 20,000(JP実カナ+ローマ字)/ postcode_validate 20,000
sft_geocode.jsonl 35,000 geocode_forward 15,000 / geocode_reverse 15,000 / display_fallback 5,000
data/changesets/*.parquet 日次 地名・郵便とも前日マスターとの差分(追記専用)

SFT合計: 765,778レコード / 18タスク(6ファイル、再計算値)

入力補助層(data/assist/)

ファイル 行数 内容
address_forms_all_countries.parquet 248 全カ国フォーム(curated 48 + 地域デフォルト 200)、欄順序 + 母国語/英語ラベル(JP prefecture=都道府県)+ postcode_required
postcode_formats.parquet 250行/248カ国 郵便番号書式(正規表現+実例)。複数書式国対応:US=ZIP5/ZIP+4、AR=4桁/CPA。HK/AE/MU/PA/QA=郵便番号なし
postcode_lookup.parquet 1,826,263 郵便番号→入力補助index:郵便番号 + 地名 + admin1/2 + 座標。JP行は漢字+カナ+ローマ字付与(romaji fill 83.0% / kana fill 84.8%)。121カ国
search_index.parquet 287,617 正規化名のソート済みindex(prefix検索用。name_norm / name_ascii_norm)
dropdown_admin1.parquet / dropdown_admin2.parquet / dropdown_cities.parquet 235,857町 プルダウン階層:国→admin1→admin2→町。admin1名欠損0.2% / admin2名欠損14.3%(DE の Bezirk なし等、構造的)
postcode_policy.parquet 248 未知郵便番号の対応ポリシー:書式既知40カ国=regex検証→前方一致候補提示、書式未確認/なし208カ国=自由入力+市町村ジオコード。いずれも入力をブロックしない

ジオコーディング層(data/geo/)

ファイル 行数 内容
geocode_places.parquet 235,857 順ジオコーディングindex:地名→lat/lon + Plus Code(OLC。実測で decode 往復誤差 <0.0001°)
geocode_postcodes.parquet 1,826,263 郵便番号重心→lat/lon + Plus Code。JP行は漢字+カナ+ローマ字付与
display_fallback_policy.parquet 249 表示困難住所の対策ポリシー:郵便番号あり国=[postcode+city → Plus Code]、なし国(HK/AE等)=[city/admin → Plus Code → lat/lon]。いずれも入力をブロックしない
scripts/reverse_geocode.py — 逆ジオコーディング helper(scipy cKDTree 最近傍)。実測:東京駅座標→京橋 0.68km / 日本橋本石町 0.80km / 皇居外苑 0.97km

精度の明示:ジオコーディング精度は町・郵便番号重心レベルです(番地・建物のrooftop精度は GeoNames に番地データがなく未対応。建物レベルは ABR(日本)等の拡張で対応予定)。

検索サーバー(案A・Typesense、実装済み)

search/typesense/ 配下。Typesense v30.2 ピン。

ファイル 内容
schema_places.json places コレクション(287,617 docs、locale:"ja"、geopoint、populationデフォルトソート)
schema_postcodes.json postcodes コレクション(~1.8M docs、121カ国、romaji/kana 任意・geopoint 任意)
build_typesense_docs.py parquet→NDJSON 変換(スキーマ検証アサーション付き)
docker-compose.yml Typesense 30.2(data volume + CORS)
import_docs.sh コレクション作成 + upsert インポート + 件数検証
README.md セットアップ手順 + オートコンプリート/ja検索/typo訂正/地理ブースト/郵便番号2層UXのクエリ例

学習済みP0モデル

  • veygrit/qwen3-0.6b-address-sft — 本データセットのSFT 765,778件(18タスク)で学習した Qwen3-0.6B LoRA r=32 adapter。train_loss 0.5984 / mean_token_accuracy 0.894 / 3,826 step・4h16m(ジョブ 6ab494b352d0dbd7f1d88df5)
  • scripts/merge_and_verify.py — アダプタをベースモデルにマージ(merge_and_unload)+ 7タスク×2プローブ×2モデルの推論検証スクリプト(実行は予算確保後)
  • メトリクス:qwen3-06b-address-sft-trackio

スキーマ(要点)

place_name_pairs.parquet — language, native, english, parent_country, place_type, geonameid, valid_from, valid_to, source(※ link/post/unlc/wkdt は除外)

han_script_pairs.parquet — script_pair(hans-hant / ja-hant), src_lang, src, dst_lang, dst, parent_country, geonameid

homograph_hanzi_places.parquet — language, hanzi, geonameid, english, parent_country, parent_admin1, type

han_reading_variants.parquet — pref_kanji, city_kanji, town_kanji, town_kana

homograph_crosscountry.parquet — language, string_name, geonameid, english, country, type, name_source(native / canonical)。疑似言語タグ(link/post/unlc/wkdt)は除外済み

languages_scripts_by_country.parquet — country, language, name_count, script, romanization_available(疑似言語タグ行は除外済み)

typo_correction_pairs.parquet — language, wrong, correct, source(synthetic-edit1+confusion。疑似言語タグ行は除外済み)

address_forms_by_country.parquet — country, fields(JSON配列), postcode_format, notes, source

address_forms_all_countries.parquet — country, lang_primary, fields(JSON配列, address1/address2含む), labels(JSON: 欄→{en, native}), postcode_required, form_variant(curated / region-default)

postcode_lookup.parquet — country, postal_code, place_name, admin1_code, admin2_code, lat, lon + JP行: pref_kanji/city_kanji/town_kanji/pref_kana/city_kana/town_kana/romaji

postcode_formats.parquet — country, format_regex, example, notes, formats(1カ国複数行=複数書式)

dropdown_cities.parquet — country, admin1_code/name/name_ascii, admin2_code/name/name_ascii, city, city_ascii, lat, lon, population

postcode_policy.parquet — country, postcode_required, allow_free_text, on_unknown, notes

geocode_places.parquet — type, geonameid, parent_country, parent_admin1, parent_admin2, name, name_ascii, lat, lon, population, olc

geocode_postcodes.parquet — country, postal_code, place_name, admin1_code, lat, lon, olc + JP行は kana/romaji 列

display_fallback_policy.parquet — country, postcode_available, fallback_chain(JSON配列), plus_code_available, notes, source

changesets — date, type/change(added/removed/updated), key/geonameid, field, old, new

履歴の保持

  1. data/changesets/ に地名・郵便の日次差分が追記され続ける(変遷を時系列で再現可能)
  2. 本リポジトリの コミット履歴 が全スナップショットの完全な履歴

データソース

  • GeoNames dump(admin階層 / alternateNames / postal)— CC BY 4.0
  • Wikidata SPARQL — CC0
  • 日本郵便 郵便番号データ(UTF-8漢字版:月次 / ローマ字版:年次)— 利用規約に従い出典明示の上で加工
  • Unihan データベース(字単位の繁簡変種)— Unicode ライセンス
  • Open Location Code(Plus Code、openlocationcode PyPI 1.0.1)— Apache-2.0
  • USPS Publication 28(米国の公式住所略称指針、public domain)
  • 配送業者の1行文字数・フォーム欄・ECフォーム欄は公開の案内・慣行に基づく初期値(curated/region-default を form_variant で区別)。業務利用前に各社の最新仕様で検証してください

毎日更新(GitHub Actions)

  1. GitHub リポジトリに scripts/ の9スクリプトと workflows/github-actions-daily.yml をコピー (workflow は .github/workflows/daily-place-names.yml に置く)
  2. Secrets に HF_TOKEN(HF 書き込みトークン)を設定
  3. cron 毎日 18:30 UTC:
  • build_places.py — 地名マスター + 差分 push
  • build_postal.py — 郵便マスター・日本ペア + 差分 push
  • build_training.py — 翻訳ペア・SFTデータ再生成 + push
  • build_han.py — 字形変換・同字異地・同字異読みレイヤー + push
  • build_disambig.py — 国境跨ぎ同名異地・typo訂正・国別言語表 + push
  • build_forms.py — フォーム欄レイヤー + push
  • build_assist.py — 全カ国フォーム+ラベル・郵便番号補助index・プルダウン・検索index・フォールバックポリシー + push
  • build_geocode.py — ジオコーディング(順/逆)・Plus Code・表示フォールバック + push
  • export_adapter_snapshot.py — (任意)veygrit-sys 国別アダプター向けスナップショット供給

手動実行:

python scripts/build_places.py --workdir /tmp/wpn --push --wikidata
python scripts/build_postal.py --workdir /tmp/wpn --push
python scripts/build_training.py --workdir /tmp/wpn --push
python scripts/build_han.py --workdir /tmp/wpn --push
python scripts/build_disambig.py --workdir /tmp/wpn --push
python scripts/build_forms.py --workdir /tmp/wpn --push
python scripts/build_assist.py --workdir /tmp/wpn --push
python scripts/build_geocode.py --workdir /tmp/wpn --push
python scripts/export_adapter_snapshot.py --workdir /tmp/wpn --push  # 任意

検索サーバーの使い方(案A・Typesense)

cd search/typesense
export TYPESENSE_API_KEY=$(openssl rand -hex 32)
docker compose up -d
python build_typesense_docs.py --out ./typesense_docs
./import_docs.sh          # places 287,617 + postcodes ~1.83M をインポート

クエリ例(オートコンプリート・日本語・typo訂正・地理ブースト・郵便番号補助)は search/typesense/README.md を参照。

Downloads last month
663