README.md

metadata

language:
  - vi
tags:
  - vietnamese
  - text-restoration
  - diacritic-restoration
  - spacing-restoration
  - nextkey
size_categories:
  - 1M<n<10M

NextKey — JDWR v1

A Vietnamese compact-writing restoration dataset for training models that restore diacritics and word boundaries.

Version

jdwr-v1 contains 881,251 aligned examples split by compacted target key to prevent overlap between train, development, and in-domain test sets.

Split	Examples	Domains
train/	577,069	7 in-domain categories
dev/	82,153	7 in-domain categories
test/in_domain/	82,078	7 in-domain categories
test/external/	159,951	the_thao external-domain evaluation

See manifest.json for per-domain counts, split policy, and overlap checks.

Schema

Each JSONL record includes compact input text, a normalized character target, a per-character whitespace-boundary target, and category/domain metadata. The current project training configuration uses the two-head character tagger over these fields.

Task scope

The dataset targets Vietnamese diacritic and spacing restoration. Brace markers are normalized by removing only the braces while retaining their content. Punctuation and capitalization restoration are out of scope for this version.

Intended use

Use for research and prototype experiments in Vietnamese compact-writing restoration. Evaluate external-domain generalization with test/external/the_thao.jsonl; do not mix it into training.