jiten-data

Pre-built dictionary assets for the Jiten app.

Downloads

• dictionary.db — download from Releases
• jlpt-words.csv — JLPT word classifications for all 22,575 common JMdict entries (Waller + audit + frequency). See jlpt-words-README.md for methodology.
• jlpt-audit-*.csv — audit override records with per-word reasoning. These are the provenance trail for corrections to Waller's JLPT classifications.
• counter-readings.csv — verified counter word readings for 264 Japanese counters × 11 numbers (1–10 + 何) = 2,904 readings. See methodology below.
• noun-counter-mappings.csv — maps 1,015 common Waller-set nouns to their appropriate counters (1,091 mappings across 88 counters). See methodology below.

License

The datasets in this repository and the pre-built dictionary databases in Releases are licensed under Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0), matching the ShareAlike terms of the upstream sources they derive from (JMdict / KANJIDIC via EDRDG, KanjiVG, WordNet, Waller/JPDB, and others). See THIRD_PARTY_NOTICES.md for full source attribution.

Counter readings methodology

264 counters were extracted from JMdict entries tagged with the "ctr" part-of-speech. For each counter, readings for numbers 1–10 and 何 were generated using phonological rules, then verified in two passes.

Generation (pass 1)

A script (scripts/build-counter-data.ts) generated candidate readings by applying standard Japanese counter phonology:

• Sokuon (っ) for 1 (いち), 6 (ろく), 8 (はち), 10 (じゅう) before eligible consonants (k-row, s-row, t-row, h→p)
• 6 sokuon restriction: only before k-row and h→p — NOT before s-row, t-row, or pure p-row loanwords
• Rendaku (h→b) for 3 (さん) and 何 (なん) before h-row
• Half-voicing exceptions (h→p instead of h→b) for specific counters like 泊(はく)

Generated readings were cross-referenced against JMdict's counted-form entries. Where JMdict had a match, that was used; where JMdict had a different reading, both were flagged for review. ~230 corrections were made in this pass.

Manual audit (pass 2)

All 2,904 readings were reviewed in 7 batches of ~30 counters each, checking:

• Sokuon rules (especially the 6-before-s/t/p restriction)
• Rendaku and half-voicing rules
• Wago (native Japanese) numeral usage for counters that require it (e.g. 人: ひとり/ふたり, 日: ついたち/ふつか)
• Irregular readings for specialized counters (mahjong, traditional units)

11 additional corrections were made in this pass (0.38% error rate from pass 1).

CSV columns

Column	Description
counter_id	JMdict sequence number
counter_kanji	Counter kanji/kana
counter_reading	Base reading of the counter
counter_gloss	English meaning
number	Number (1–10 or 何)
number_kanji	Kanji for the number
combined_kanji	Number + counter kanji
generated_reading	Reading from phonological rules
jmdict_readings	JMdict counted-form readings (if any)
source	jmdict_match, jmdict_mismatch, or generated
notes	Reasoning for mismatches or corrections
verified_reading	Final verified reading

Noun–counter mappings methodology

1,015 common nouns from the Waller set (5,303 total Waller nouns) were mapped to their standard Japanese counters.

Generation

The 5,303 Waller-set nouns were split into 6 batches and each batch was classified against the full list of 264 counters. The prompt asked for only standard, natural counters — not generic fallbacks like つ.

This produced 1,279 candidate mappings.

Manual audit

All 1,279 mappings were reviewed in 12 batches of ~100, checking for:

• Wrong counter type: e.g. 台 assigned to non-machines (stores, factories), 本 for non-cylindrical things (paintings), 枚 for non-flat things (terms)
• Non-countable nouns: abstract concepts (理解, 裏), na-adjectives (幼稚), adverbial expressions (一人一人, 種々)
• Self-referential mappings: counter expressions mapped to themselves (一人→人, 二人→人, 万人→人)
• Missing counters: entries referencing counter IDs not in the dataset
• Wrong sense: words with multiple meanings mapped using the wrong sense (品 as "elegance" → 品 counter)
• Overuse patterns: 台 for anything with a surface, 品 for non-food items, 人 for pronouns/collectives

188 mappings were removed (14.7%) and ~18 were corrected to different counters.

CSV columns

Column	Description
jmdict_id	JMdict sequence number of the noun
counter_id	JMdict sequence number of the counter