NagaYu/rebar-structure

🧱 Rebar Structure

A corpus for restoring the heading hierarchy of Japanese documents from flat text and for evaluating structure-aware chunking. Each record is a flattened document, its gold heading tree (positions + depths), and a damaged variant simulating PDF/text extraction.

• Code: https://github.com/NagaYu/rebar
• Model: https://huggingface.co/NagaYu/rebar-heading-classifier
• Demo (Space): https://huggingface.co/spaces/NagaYu/rebar

Why it exists

The same numbering marker means different depths in different documents (1. is top-level in a memo, second-level under 第3章), so hierarchy must be inferred at the document level. This dataset is built to measure exactly that, and — crucially — to measure generalization to numbering schemes held out of training (the hard split).

Splits

Split	What	Note
train	seen numbering schemes	公用文 / 法令 / 規程 / simple-numeric chains
test	seen schemes, different documents	document-level split, no leakage
hard	UNSEEN numbering schemes	held-out chains, incl. one that contradicts the canonical prior (paren shallower than arabic-dot)

Fields

Field	Type	Description
doc_id	string	unique id
text	string	flattened plain text; visible numbering preserved
damaged_text	string	text-extraction-damaged variant (mid-line breaks, header/footer, spacing, ruby, broken tables, page numbers, width jitter)
headings	string (JSON)	gold list of {line, depth, kind, value, marker, title}
n_headings	int	number of gold headings
max_depth	int	deepest gold level
scheme_name	string	numbering scheme id
source	string	synthetic or a local source path
license	string	per-document license

import json
from datasets import load_dataset
ds = load_dataset("NagaYu/rebar-structure")
ex = ds["hard"][0]
headings = json.loads(ex["headings"])   # gold hierarchy
print(ex["scheme_name"], len(headings), "headings, max depth", ex["max_depth"])

Provenance & license

The published corpus is synthetic and CC0, generated by rebar.synth with fully controlled gold trees — so every benchmark reproduces from the repository alone, with no redistribution concerns.

Real documents are not redistributed here. To include licensed real Markdown/HTML, use scripts/build_dataset.py with --redact, which ships only structure labels + minimal fragments (never full source bodies). Always verify a source's redistribution terms before adding it.

Intended use & limitations

Built for hierarchy-restoration and chunking research. Synthetic text is regulation-flavoured but not real regulations; absolute numbers transfer best as relative comparisons between methods. Heavily damaged, unseen-scheme documents are the hardest case and are reported honestly in the benchmarks.