doc-split-benchmark / README.md
hung-k-nguyen's picture
Release snapshot
eb65bfd
|
Raw
History Blame Contribute Delete
3.31 kB
metadata
license: cc-by-sa-4.0
pretty_name: Doc-Split Benchmark (page-stream segmentation)
task_categories:
  - image-classification
  - text-classification
language:
  - multilingual
tags:
  - page-stream-segmentation
  - document-boundary-detection
  - document-ai
  - benchmark
size_categories:
  - 1K<n<10K
configs:
  - config_name: our200
    data_files:
      - split: test
        path: our200/test-*.parquet

Doc-Split Benchmark

The evaluation slice for page-stream segmentation — the exact set behind the leaderboard and the cloud-VLM comparison. Self-contained (page images embedded), with a reference scorer so results are reproducible. This is the benchmark, not the training corpus (which stays private).

Cuts

Cut Where Notes
our-200 this repo (our200 config) 200 license-clean synthetic-concat streams, ~78% boundary rate
OpenPSS short / long openpss-mirror Dutch; sparse- & dense-boundary regimes
TABME++ / Tobacco800 upstream (not redistributed) reported for reference; licenses/gating prevent re-hosting

Load

from datasets import load_dataset
ds = load_dataset("nutrientdocs/doc-split-benchmark", "our200", split="test")
# one row = one page; group by stream_id, order by position.

Schema

Field Type Meaning
stream_id string groups pages into one ordered stream
position int page index within the stream (0-based)
boundary int 1 = starts a new document (position 0 always 1)
page_text string page OCR text (may be empty)
image image page image (512px)
source string provenance

How to score

Metric: boundary page-F1 (internal positions, page 0 forced) + Cohen's κ, WindowDiff, Pk. The reference scorer ships in this repo:

python score.py --pred your_preds.json      # preds: {stream_id: [0/1 per page]}

score.py uses docsplit.eval.metrics.aggregate (bundled). κ is reported because a degenerate predict-none/all can look fine on F1 but scores κ ≈ 0 on the sparse cuts. See SUBMISSION.md to add a row to the leaderboard.

manifest.json records provenance (per-source stream/page counts, boundary rates, seed).

License & attribution

CC-BY-SA-4.0 for the our-200 packaging (Nutrient-generated synthetic concatenations). OpenPSS cuts are in openpss-mirror under their upstream terms. TABME++/Tobacco800 are referenced, not redistributed.

About the author

This project is maintained and funded by Nutrient - The #1 PDF SDK library for viewing, editing, eSigning, and more.