POTATR-v1.1-Pub

Paper: POTATR: A Lightweight Image-to-Graph Model for Page-Level Table Extraction by Brandon Smock, Libin Liang, Maxim Sokolov, Amrit Ramesh, Valerie Faucon-Morin, Tayyibah Khanam, and Maury Courtland (DAS 2026)

POTATR (Page-Object Table Transformer) is a transformer-based model for recognizing tables on document pages. From a page image it detects each table, the objects that make up its structure (rows, columns, column headers, projected row headers, spanning cells), as well as its surrounding context (captions and footers). It predicts a graph that links each table to its objects. With the page's words, the predicted structure is turned into table cells with their text.

These are the weights of POTATR v1.1, trained on the Single Pages and Full Documents collections of PubTables-v2. The code to run, evaluate and train the model is at https://github.com/kensho-technologies/POTATR.

Files

• POTATR-v1.1-Pub.safetensors: the model weights.
• potatr_v1.json: the model configuration that goes with them (also configs/model/potatr_v1.json in the code repository).

Usage

Install the code from the repository, then run the model on a directory of page images and a directory of the pages' words:

potatr-infer \
    --config_file potatr_v1.json \
    --model_load_path POTATR-v1.1-Pub.safetensors \
    --images_dir pages/images \
    --words_dir pages/words \
    --output_dir output/tables

See the repository's README for the input formats, the output, and evaluation on PubTables-v2.

Training

POTATR v1.1 was trained in two stages on 8 NVIDIA T4 GPUs, starting from TATR-v1.1-Pub (Table Transformer v1.1 trained on PubTables-1M):

1. 97 epochs on the Single Pages training split (this is POTATR v1.0);
2. 35 more epochs on the Single Pages training split, a second pass over its pages with long tables, and 2,000 random pages without tables from the Full Documents training split each epoch.

The repository has the configs and commands for both stages.

License

The weights are released under the MIT License. They were trained on PubTables-v2, which is licensed under CDLA-Permissive-2.0, starting from TATR-v1.1-Pub (MIT License).

Citation

@inproceedings{smock2026potatr,
  title={{POTATR}: A Lightweight Image-to-Graph Model for Page-Level Table Extraction},
  author={Smock, Brandon and Liang, Libin and Sokolov, Maxim and Ramesh, Amrit and Faucon-Morin, Valerie and Khanam, Tayyibah and Courtland, Maury},
  booktitle={International Workshop on Document Analysis Systems},
  pages={3--17},
  year={2026},
  organization={Springer}
}

If you use the PubTables-v2 dataset, please also cite:

@article{smock2025pubtables,
  title={Pub{T}ables-v2: A new large-scale dataset for full-page and multi-page table extraction},
  author={Smock, Brandon and Faucon-Morin, Valerie and Sokolov, Max and Liang, Libin and Khanam, Tayyibah and Ramesh, Amrit and Courtland, Maury},
  journal={arXiv preprint arXiv:2512.10888},
  year={2025}
}