---
license: cc-by-nc-sa-4.0
language:
- en
pretty_name: SciTSR-CC-BY-NC-SA
size_categories:
- n<1K
task_categories:
- image-to-text
- object-detection
tags:
- table-structure-recognition
- document-understanding
- scientific-tables
- pdf
- non-commercial
---

# SciTSR-CC-BY-NC-SA

A license-filtered subset of [SciTSR](https://github.com/Academic-Labeling/SciTSR), a large-scale table structure recognition dataset of scientific tables extracted from arXiv LaTeX source files.

This subset contains tables whose source papers are compatible with a **CC-BY-NC-SA 4.0** open-weight model release — covering public domain, CC-BY, CC-BY-NC, and CC-BY-NC-SA licensed papers. The dataset itself is released under CC-BY-NC-SA 4.0.

## Dataset Details

| Split | Tables | Papers |
|-------|-------:|-------:|
| train | 697 | — |
| test  | 192 | — |
| **total** | **889** | **470** |

Source paper licenses present:

| License | Tables | Notes |
|---------|-------:|-------|
| `CC-BY` | 444 | Attribution required |
| `CC-BY-NC-SA` | 337 | NC + share-alike |
| `CC0` | 85 | Public domain |
| `other:http://creativecommons.org/licenses/publicdomain/` | 23 | Old CC public domain dedication |

## Why This Subset Exists

The full SciTSR dataset (15,000 tables) was crawled from arXiv without license filtering. ~90% of those papers use the arXiv non-exclusive license, which retains full author copyright. This subset was produced by querying the arXiv OAI-PMH API for the license of each source paper and retaining only those compatible with a non-commercial open-weight release.

**License compatibility logic**: CC-BY-NC-SA on model weights requires that training data allows NC use and that any SA clause matches the output license. CC-BY-SA was explicitly excluded because its SA clause conflicts with adding an NC restriction. This analysis assumes model weights are not derivative works of training data — a legally unsettled question. Users should conduct their own legal review for their specific use case.

For unrestricted commercial use with no attribution requirements, see [SciTSR-PD](https://huggingface.co/datasets/rootsautomation/SciTSR-pd).

## Dataset Structure

Each row represents one table extracted from a scientific PDF. The `paper_license` column records the source paper's license so downstream users can apply their own filtering.

| Column | Type | Description |
|--------|------|-------------|
| `table_id` | `string` | Unique identifier, format `{arxiv_id}v{version}.{table_index}` |
| `paper_id` | `string` | arXiv paper ID |
| `paper_title` | `string` | Paper title from arXiv metadata |
| `paper_authors` | `list[string]` | Author names from arXiv metadata |
| `paper_license` | `string` | License of the source paper |
| `split` | `string` | `train` or `test` |
| `is_comp` | `bool` | Whether this table is in the SciTSR-COMP subset (tables with at least one spanning cell) |
| `image` | `Image` | PNG render of the table (150 DPI) |
| `pdf` | `binary` | Raw PDF of the isolated table |
| `chunks` | `list[dict]` | Pre-extracted text spans with bounding box coordinates `{text, x1, x2, y1, y2}` (PDF coordinate space, bottom-left origin) |
| `cells` | `list[dict]` | Structure annotation: `{id, tex, content, start_row, end_row, start_col, end_col}` |
| `relations` | `list[dict]` | Chunk adjacency labels `{chunk_id_1, chunk_id_2, relation, num_blank}` where `relation=1` is horizontal and `relation=2` is vertical. Empty for test rows. |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("rootsautomation/SciTSR-cc-by-nc-sa")

# Iterate train split
for row in ds["train"]:
    image   = row["image"]          # PIL Image
    cells   = row["cells"]          # list of cell dicts
    chunks  = row["chunks"]         # list of chunk dicts with positions
    print(row["table_id"], row["paper_title"], row["paper_license"])

# Filter to only public domain if needed
pd_only = ds.filter(lambda x: "CC0" in x["paper_license"] or "publicdomain" in x["paper_license"])
```

## Important Caveats

- **Annotation quality**: Structure annotations were generated automatically from LaTeX source by the original SciTSR pipeline. Simple grid tables are generally reliable. Tables with spanning cells (`is_comp=True`) have a higher rate of annotation errors, particularly in spanning cell coordinates. Treat annotations as noisy weak supervision rather than ground truth.
- **Attribution**: CC-BY, CC-BY-NC, and CC-BY-NC-SA papers require attribution to original authors. `paper_title`, `paper_authors`, and `paper_id` (linking to `arxiv.org/abs/{paper_id}`) are provided for this purpose.
- **Chunk coordinates**: The `chunks` field was pre-processed by [TabbyCDF](https://github.com/cellsrg/tabbypdf/) and may contain noise.
- **Relations**: Available for train split only; empty list for test.

## Source & Citation

This dataset is a license-filtered derivative of SciTSR. If you use this dataset, please cite the original work:

```bibtex
@article{chi2019complicated,
  title={Complicated Table Structure Recognition},
  author={Chi, Zewen and Huang, Heyan and Xu, Heng-Da and Yu, Houjin and Yin, Wanxuan and Mao, Xian-Ling},
  journal={arXiv preprint arXiv:1908.04729},
  year={2019}
}
```