einkpub — model weights for einkpdf

NNEF exports of the two models einkpdf uses to turn research PDFs into EPUBs sized for a 7" e-ink reader. Both run under tract on CPU — no GPU, and no Python at inference time.

A page is rasterized, the layout model finds the labelled regions and their reading order, and the transcriber turns the page into structured markdown. The layout half exists because the transcriber emits no coordinates at all, so there would otherwise be nothing to crop a figure from.

⚠️ This repository holds two models under two different licences

HuggingFace has one licence field per repository, so this one is tagged other. The licence depends on which files you use.

Path	Model	Licence
/	NuExtract3 transcriber	Apache-2.0 — see LICENSE, NOTICE
/layout	surya layout predictor	CC-BY-NC-SA-4.0 — see layout/LICENSE, layout/NOTICE

If you take only the transcriber, Apache-2.0 governs and there is no non-commercial restriction. If you take anything under layout/, the CC-BY-NC-SA-4.0 terms apply to those files, including the non-commercial and share-alike conditions.

surya's author waives the non-commercial term for organizations under $5M USD in gross revenue over the most recent 12 months and under $5M in lifetime VC/angel funding, that are not competitive with the Datalab API. That waiver is theirs to grant, not this repository's to pass on — commercial licensing is available from Datalab.

Files

Transcriber — Apache-2.0

File	Size	Purpose
vision.nnef.tgz	1.3 GB	Vision tower
decoder.nnef.tgz	2.3 GB	Text decoder
tokenizer.json	20 MB	HuggingFace tokenizer
einkpdf.json	4 KB	Runtime config consumed by einkpdf

Layout — CC-BY-NC-SA-4.0

File	Size	Purpose
layout/layout_predictor_encoder.nnef.tgz	299 MB	Vision encoder
layout/layout_predictor_decoder_prefill.nnef.tgz	219 MB	Decoder, prefill pass
layout/layout_predictor_decoder_iter.nnef.tgz	199 MB	Decoder, iterative pass
layout/layout_config.json	1 KB	Slicer and preprocessor config
layout/model_config.json	6 KB	Decoder token config

Changes from the base models

Transcriber, from numind/NuExtract3:

• Converted from safetensors to NNEF for tract.
• Weights at f16; the decoder additionally quantized to Q4_0, cutting peak RSS from ~21 GB to ~6 GB and running ~1.76x faster on CPU.
• No fine-tuning.

Layout, from surya (checkpoint 2025_02_18):

• Converted from PyTorch to NNEF, split into encoder, decoder-prefill and decoder-iterative graphs.
• No fine-tuning.

Neither conversion changes model architecture, and both are intended to match their originals. The layout preprocessor is checked against golden arrays from surya's Python implementation.

A note on operator names

The layout graphs were exported against tract 0.21 and still name tract_onnx_is_nan, which tract 0.23 moved into core as tract_core_is_nan. einkpdf re-registers the old name at load time rather than requiring a re-export; anything else loading these graphs on tract ≥ 0.23 will need to do the same.