NagaYu/rebar-heading-classifier

🧱 Rebar — heading classifier

Rebar restores the heading hierarchy of Japanese documents from flattened plain text (第1章 → 1. → (1) → ア), then chunks along that hierarchy so retrieval keeps articles, tables and FAQs intact. This repository hosts the heading-line detector — the one learned component. It is deliberately tiny: a logistic regression over 13 engineered features, sigmoid(x·w + b), exported to ONNX (~0.6 KB) and GGUF (~0.7 KB).

• 👉 Code & docs: https://github.com/NagaYu/rebar
• 👉 Live demo (Space): https://huggingface.co/spaces/NagaYu/rebar
• 👉 Dataset: https://huggingface.co/datasets/NagaYu/rebar-structure

Rebar is not a large multimodal PDF parser and does no layout analysis. It is the stage after text extraction. Its edge is local, cheap, layout-independent.

Why a tiny model works

The hard part isn't a big network — it's the document-level inference around it. The same marker means different depths in different documents (1. is top-level in a memo, second-level under 第3章), so depth is inferred jointly for the whole document from its nesting evidence, using the Japanese public-document / statutory chains as a prior. The classifier here only answers "is this line a heading?"; scheme inference, tree decoding and chunking are deterministic algorithms in the package.

Results (synthetic benchmark)

Set	Heading F1	Depth accuracy	Tree agreement
Test (seen schemes, clean)	0.997	1.000	1.000
Hard (UNSEEN schemes, clean)	0.995	1.000	1.000
Test (damaged)	0.53	0.81	0.59

Downstream, chunking along the restored tree drops known failure modes to 0% (article↔proviso, table↔title, FAQ Q↔A) and lifts complete-answer recall@1 (e.g. tables 96.6% vs. ~64% for fixed-length).

Usage

Highest level — the full pipeline (recommended):

pip install "git+https://github.com/NagaYu/rebar"

from rebar import Rebar
text = open("doc.txt", encoding="utf-8").read()
print(Rebar().outline(text).outline())      # restored hierarchy
chunks = Rebar().chunk(text, max_len=800)    # structure-aware chunks

Just the ONNX classifier (line → P(heading)):

import numpy as np, onnxruntime as ort
from rebar.model import line_features           # 13-dim feature extractor
sess = ort.InferenceSession("rebar_heading_clf.onnx")
x = line_features("第3条 目的").astype(np.float32)[None, :]
print(sess.run(None, {"features": x})[0])       # ~1.0 => heading

Files

File	What
heading_clf.json	weights + feature names (loaded by the Python package)
rebar_heading_clf.onnx	portable graph, verified to match numpy at export
rebar_heading_clf.gguf	spec-compliant GGUF weight container (arch rebar-logreg) — an honest weight container, not a llama.cpp LLM

Features (13)

has_marker, marker_rank_norm, is_appendix, is_short, title_len_norm, ends_sentence, has_tab, looks_page_number, digit_ratio, punct_ratio, spaced_ratio, layout_indent, layout_heading_font. The two layout_* slots are optional auxiliary inputs — zero when no layout is available (Rebar runs without layout).

License

Apache-2.0. Training data is the CC0 synthetic corpus in NagaYu/rebar-structure.