Dataset Viewer
Auto-converted to Parquet Duplicate
source_id
stringclasses
5 values
upstream
stringclasses
4 values
source_url
stringclasses
5 values
archive_bytes
int64
11M
11.2B
archive_sha256
stringclasses
5 values
recorded_terms
stringclasses
4 values
ai2d
https://prior.allenai.org/projects/diagram-understanding
https://ai2-website.s3.amazonaws.com/data/ai2d-all.zip
990,965,374
1a6b77eebb8b7dbdf76a0ba6ca76c2f97ce8f81d8ee33b06593aa722e54c4786
AI2D research-only license (non-commercial; no redistribution)
chartqa
https://github.com/vis-nlp/ChartQA
https://huggingface.co/datasets/ahmed-masry/ChartQA/resolve/af8b6f5c08c95085271561c2a3f9d15f2b5a9031/ChartQA%20Dataset.zip?download=true
875,370,872
1bf310e5a51101681495c4a24f4f29d22c4f70b52df24e2e4feb0d79cae3c160
GPL-3.0
docci_annotations
https://google.github.io/docci/
https://storage.googleapis.com/docci/data/docci_descriptions.jsonlines
11,000,214
c9df4819963883af35ddd2cf257949892fd8c6d88b33a012094352df60719800
CC-BY-4.0
docci_images
https://google.github.io/docci/
https://storage.googleapis.com/docci/data/docci_images.tar.gz
7,592,938,768
c1b1aee00856757d71cfe2cc6ab641089284276a791bfb82146b4742c59a0a1d
CC-BY-4.0
pubtabnet
https://github.com/ibm-aur-nlp/PubTabNet
https://huggingface.co/datasets/ajimeno/PubTabNet/resolve/06963b1af16203f4633c718e2c50109eb57ed658/pubtabnet.tar.gz?download=true
11,244,059,914
90c55e733c85c98edf6d350b77f1e4c23767577555fc932e44ad723674de8d3e
CDLA-Sharing-1.0

Opsis V1 training data provenance

This repository records the local training-data snapshot used for opsis-v1-nano-g3. It does not contain a loadable training split, images, annotations, or generated targets. The mixed-source corpus includes AI2D, whose bundled license forbids distribution to third parties and public release of modified data. The full corpus therefore remains local.

The model card describes the checkpoint, evaluation, and CPU runtime. This page identifies the training inputs and their source terms without implying that the complete dataset can be downloaded here.

The Dataset Viewer shows five source archive metadata records: upstream URLs, archive sizes, checksums, and recorded terms. Its source_archives split is not a training or validation split.

Snapshot

Item Recorded value
Training rows 100,000
Validation rows 1,000
Distinct training image paths 96,875
Repeated training image rows 3,125
Distinct image files across train and validation 97,875
Local image bytes across both splits 20,691,179,759
Visually reviewed gold benchmark cases 600
Gold cases overlapping training, per curation report 0

The training mix comprises 31,250 tables, 18,750 charts, 18,750 diagrams, and 31,250 ordinary images. Seventy-five thousand examples are clean images and 25,000 are PDF-style crops. The corrected chart mix includes 12,500 ChartQA and 6,250 synthetic chart examples.

PROVENANCE.md records the SHA-256 of the corrected train and validation manifests, curation reports, curator code, source archives, and model weights. It also records a SHA-256 fingerprint over the bytes of every local image file used by the two splits, in sorted relative-path order. These hashes identify the local snapshot; they do not make the restricted files available from this repository.

The row manifests currently contain machine-local absolute paths from the original training workspace. The images were verified at their corresponding locations in the current Opsis workspace. The public provenance record contains no local user path or row-level content.

Source mixture

Source Role Recorded terms
PubTabNet Tables CDLA Sharing 1.0
ChartQA Charts GPL 3.0
AI2D Diagrams Research-only; no third-party redistribution under the bundled license
DOCCI Ordinary images CC BY 4.0
Locally generated charts and diagrams Synthetic augmentation Project-generated; source dependencies remain subject to their terms

This list summarizes the source metadata recorded during curation. It is not a single license grant for the mixture or the released model. Consult each upstream source before using or redistributing any part of the training corpus.

What this record verifies

The hashes can confirm that a local copy has the same row manifests, source archives, and image bytes as the recorded Opsis G3 snapshot. They do not provide the training images, make restricted material redistributable, or by themselves reproduce the checkpoint. Training also requires the base checkpoint, code, configuration, and compute environment described on the model card.

Downloads last month
45