doc-split-v2 / README.md
hung-k-nguyen's picture
Release snapshot
61cd31a
|
Raw
History Blame Contribute Delete
4 kB
metadata
license: other
license_name: nutrient-commercial
pipeline_tag: image-text-to-text
language:
  - multilingual
tags:
  - page-stream-segmentation
  - document-boundary-detection
  - document-ai
  - document-splitting
datasets:
  - nutrientdocs/doc-split-benchmark

doc-split-v2 · commercial

One model that splits any document stream — short or long, English or not, forms or prose. The high-accuracy flagship for page-stream segmentation: hand it a merged/scanned PDF and it marks where each new document begins. Weights are commercial (not downloadable here); this page is a spec + scorecard.

It is the document specialist in a field of generalists — cloud VLMs and OpenPSS's own single-slice specialists each fall down somewhere; this one does short and long streams with one model. For the open-weight companion, see doc-split-v1.

Results — boundary F1 (κ)

This model vs the open doc-split-v1, the strongest cloud VLM, and prior work (bold = best releasable single model).

Cut doc-split-v2 doc-split-v1 (open) best cloud VLM OpenPSS specialist
OpenPSS-short (sparse) 0.652 (.60) 0.585 0.598 0.76 (short-spec)
OpenPSS-long 0.891 (.86) 0.859 0.244 0.83 (long-spec)
our-200 (synthetic) 0.944 (.79) 0.936 0.942
TABME++ test 0.943 (.91) 0.704
Tobacco800 test 0.969 (.93) 0.820
val (real-doc) 0.917 (.86) 0.918

One balanced model, not two specialists. OpenPSS needs a separate short- and long-specialist (each craters on the other slice); the flagship does both with one model, and its OpenPSS-long (0.891) tops even OpenPSS's own long-specialist (0.83) and every cloud VLM (best 0.244). It also dominates the modern TABME++/Tobacco800 benchmarks. See the leaderboard for the full field.

Intended use & limits

  • Use it for: segmenting a stream of page images (a merged/scanned PDF) into its constituent documents — short or long, English or not, forms or prose. One model handles both sparse and dense boundary regimes; the embedded text layer sharpens boundaries when present, and scanned pages fall back to vision.
  • Limits: optimized for document page streams; confidence is calibrated on our-domain data, so very out-of-distribution scans are approximate. Boundaries only (not document type).

License & data

The model weights are offered under a commercial Nutrient license — deployed on-prem, so your documents never leave your infrastructure. The training set is not redistributed; evaluation runs on the held-out doc-split-benchmark.

📩 Get access

doc-split-v2 is commercial and its weights are not downloadable here. To run it on-prem — one model for any stream, calibrated, private — contact Nutrient: nutrient.io/contact-sales.

About the author

This project is maintained and funded by Nutrient - The deterministic document infrastructure enterprises run their highest-stakes workflows on: replayable output, clear exceptions, and full audit trails on the messy, regulated documents where AI alone breaks.