README.md

---
license: other
license_name: cc-by-4.0-compilation-with-upstream-terms
language:
- ja
size_categories:
- 1M<n<10M
tags:
- japanese
- ime
- dictionary
configs:
- config_name: conversion
  data_files:
  - split: train
    path: entries.tsv
- config_name: prediction
  data_files:
  - split: train
    path: predictions.tsv
---
# IME Dictionary Core

Japanese IME dictionary pack. The SQLite database and TSV files contain the
same conversion and prediction records. See `NOTICE.md` before redistribution.

- Compilation license: `CC-BY-4.0`
- SQLite tables: `entries`, `entry_sources`, `predictions`, `prediction_sources`
- TSV exports: `entries.tsv`, `predictions.tsv`

## Kaomoji flag

The SQLite `kaomoji` column (0/1) marks symbol-structured emoticons; filter
with `WHERE NOT kaomoji` for a face-free dictionary. A candidate is flagged
when it matches an ASCII emoticon (`:-)`, `orz`, `T_T`), looks structurally
like a kaomoji (bracket / face-mark / arm-mark heuristics, <= 48 chars), or
is sourced solely from the project's kaomoji dictionaries. Unicode emoji are
legitimate conversions and are NOT flagged.

## Cost calibration

Costs are on Mozc's word-cost scale; layers merged in order, the first
provider of a (reading, candidate) wins. Project-original dictionaries were
quantile-mapped onto Mozc via their shared entries, Mozc was added
unchanged, then each remaining layer was isotonic-calibrated on its overlap
with the pack and offset by a fixed penalty, adding only novel entries:
`ipadic` +5000, `sudachi` +6000, `skk` (core subset) +5000. Pre-normalization: sudachi `clamp(7000 + cost/4, 6000,
16000)`; original emoji files `18000 + clamp(cost, 0, 10000)/2`; other
original files clamped to per-file floors (default 4500) and a 30000
ceiling.

## Related repository

For the larger GPL-licensed vocabulary, see [ime-dictionary-extended-gpl](https://huggingface.co/datasets/fa0311/ime-dictionary-extended-gpl).