slanext-wired-raw-datasets

Converted raw datasets for SLANeXt_wired table structure finetuning. Each split contains PNG table images and JSONL annotations with HTML structure tokens and cell bounding boxes.

Summary

• Total samples: 61,359
• Total images: 63,929
• On-disk size (local): 7.0 GB
• Generated: 2026-06-21 15:46 UTC
• Includes source archives: no (images + annotations only)

Dataset mix

Dataset	Samples	Images	Size	Upstream source
fintabnet	20,000	20,000	3.2 GB	apoidea/fintabnet-html
pubtables	40,000	40,000	1.0 GB	apoidea/pubtabnet-html
scitsr	889	939	27.9 MB	rootsautomation/SciTSR-cc-by-nc-sa
icdar	250	990	317.6 MB	bsmock/ICDAR-2013-Table-Competition-Corrected
marmot	220	2,000	2.4 GB	https://www.icst.pku.edu.cn/cpdp/docs/20190424190300041510.zip

Layout

<dataset>/
  annotations.jsonl   # one JSON object per table sample
  hf_source.json      # download provenance
  images/             # PNG renders

Annotation format

Each JSONL row contains:

• filename: relative path under the dataset folder
• html.structure.tokens: table HTML token sequence
• html.cell[]: cell text tokens + bbox [x1,y1,x2,y2,...]

Example:

{
  "filename": "images/fintabnet_000000.png",
  "html": {
    "structure": {"tokens": ["<table>", "<tr>", "<td>", "..."]},
    "cell": [{"tokens": ["Year"], "bbox": [1, 1, 330, 56]}]
  }
}

Usage

pip install huggingface_hub
huggingface-cli download AvoCahDoe/slanext-wired-raw-datasets --repo-type dataset --local-dir ./data/raw

Then run the finetune pipeline from TableDetectionRec/finetune:

python scripts/02_convert_all.py
python scripts/04_train_staged.py --device gpu:0

License notes

Upstream licenses apply per source dataset (FinTabNet, PubTabNet, SciTSR, ICDAR 2013, MARMOT). Review each upstream repository before redistribution or commercial use.