orgrctera/kleister_charity_information_extraction

Kleister Charity — Information Extraction

Dataset description and background

Kleister Charity is one of two corpora introduced in Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts (ICDAR 2021). It targets Key Information Extraction (KIE) from long, formally structured English documents that mix scanned pages and born-digital PDFs, with complex visual layout (tables, multi-column text, headers, etc.).

The Charity split is built from annual financial and activity reports filed by UK charity organizations. At the scale reported in the original paper, the full benchmark comprises on the order of ~2,788 reports, ~61k unique pages, and ~21k annotated entities to extract—making it a demanding testbed for models that must combine textual understanding with layout and document structure.

Systems are evaluated on recovering typed fields (e.g. charity identifiers, addresses, financial figures, reporting dates) that appear in heterogeneous positions and formats across documents.

Primary references (see References):

• arXiv:2105.05796 — full paper with task definition, statistics, and baselines (Flair, BERT, RoBERTa, LayoutLM, LAMBERT, etc.).
• Official GitHub release — applicaai/kleister-charity — data and tooling associated with the Charity corpus.

This Hugging Face repository is a CTERA-packaged export of the information extraction task for sub-benchmark charity, derived from evaluation traces (e.g. Langfuse). It is intended for benchmarking and RAG / document-AI pipelines in a columnar, Parquet-friendly layout.

Task: Information Extraction (Key Information Extraction)

This dataset supports Information Extraction / Key Information Extraction (KIE):

• Input: A document identifier (input.filename, typically a hashed PDF name in this export) together with the ordered list of fields to extract (input.keys).
• Target: A structured record—here a JSON object mapping each field name to its string value as extracted from the document (expected_output).

In the original Kleister setting, strong performance typically requires using both text and layout (e.g. layout-aware transformers). The paper reports strong baseline F1 on Kleister Charity (best reported system ~83.57% F1 in the paper’s experiments), illustrating that the task remains non-trivial even for modern models.

Fields to extract (schema)

Each row lists the same logical key set in input.keys (space-separated). The benchmark targets attributes such as:

Logical field	Role (high level)
charity_name	Registered name of the charity
charity_number	Official charity registration number
report_date	Reporting period / report date
address__street_line, address__post_town, address__postcode	Registered address components
income_annually_in_british_pounds, spending_annually_in_british_pounds	Annual income and spending (GBP)

Exact key names and normalization follow the Kleister Charity annotation convention (see the paper and original release).

Data fields

Parquet splits expose the following columns:

Column	Description
id	Stable row identifier (UUID string).
input.filename	Document key (hashed PDF filename) for the source report in this export.
input.keys	Space-separated list of field names the model must fill.
expected_output	JSON string: object mapping field names to gold string values on train; may be empty on test when labels are withheld for evaluation.
metadata.item	JSON string with auxiliary info (e.g. repeated keys, nested expected, filename echoes).

Splits

Split	Rows (this HF dataset)
train	1,729
test	609
Total	2,338

Note: In this export, test rows may carry empty expected_output while evaluation is performed against held-out or external references—align your metric script with your organization’s Kleister evaluation protocol.

Examples

Train (gold expected_output present)

{
  "id": "572e4a16-041f-45b7-8c42-8729d144458c",
  "input.filename": "1c86b1fb218c77d353bb75074f97edb6.pdf",
  "input.keys": "address__post_town address__postcode address__street_line charity_name charity_number income_annually_in_british_pounds report_date spending_annually_in_british_pounds",
  "expected_output": "{\"report_date\": \"2015-03-31\", \"charity_name\": \"Discrimination_Law_Association\", \"charity_number\": \"1124892\", \"address__postcode\": \"N6_5HN\", \"address__post_town\": \"LONDON\", \"address__street_line\": \"33_CROMWELL_AVENUE\", \"income_annually_in_british_pounds\": \"29312.00\", \"spending_annually_in_british_pounds\": \"31172.00\"}",
  "metadata.item": "{\"keys\": \"address__post_town address__postcode address__street_line charity_name charity_number income_annually_in_british_pounds report_date spending_annually_in_british_pounds\", \"expected\": { ... }, \"filename\": \"1c86b1fb218c77d353bb75074f97edb6.pdf\"}"
}

(The metadata.item value is abbreviated with ... in the example; the stored row contains the full JSON.)

Test (expected_output may be empty)

{
  "id": "5c3ff34c-0927-4e98-a087-ce0934125b77",
  "input.filename": "1bf755592ecfd20cec8adeab143eddcc.pdf",
  "input.keys": "address__post_town address__postcode address__street_line charity_name charity_number income_annually_in_british_pounds report_date spending_annually_in_british_pounds",
  "expected_output": "",
  "metadata.item": "{\"keys\": \"address__post_town address__postcode address__street_line charity_name charity_number income_annually_in_british_pounds report_date spending_annually_in_british_pounds\", \"expected\": null, \"filename\": \"1bf755592ecfd20cec8adeab143eddcc.pdf\"}"
}

References

Paper (abstract)

Tomasz Stanisławek, Filip Graliński, Anna Wróblewska, Dawid Lipiński, Agnieszka Kaliska, Paulina Rosalska, Bartosz Topolski, Przemysław Biecek.
Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts.
Presented at ICDAR 2021. Preprint: arXiv:2105.05796.

The relevance of the Key Information Extraction (KIE) task is increasingly important in natural language processing problems. But there are still only a few well-defined problems that serve as benchmarks for solutions in this area. To bridge this gap, we introduce two new datasets (Kleister NDA and Kleister Charity). They involve a mix of scanned and born-digital long formal English-language documents. In these datasets, an NLP system is expected to find or infer various types of entities by employing both textual and structural layout features. The Kleister Charity dataset consists of 2,788 annual financial reports of charity organizations, with 61,643 unique pages and 21,612 entities to extract. The Kleister NDA dataset has 540 Non-disclosure Agreements, with 3,229 unique pages and 2,160 entities to extract. We provide several state-of-the-art baseline systems from the KIE domain (Flair, BERT, RoBERTa, LayoutLM, LAMBERT), which show that our datasets pose a strong challenge to existing models. The best model achieved an 81.77% and an 83.57% F1-score on respectively the Kleister NDA and the Kleister Charity datasets. We share the datasets to encourage progress on more in-depth and complex information extraction tasks.

Additional links

• PDF: https://arxiv.org/pdf/2105.05796
• Springer chapter DOI: 10.1007/978-3-030-86549-8_36
• Author page (summary): Kleister — Przemysław Biecek

BibTeX (arXiv)

@article{stanislawek2021kleister,
  title={Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts},
  author={Stanis{\\l}awek, Tomasz and Grali{\\'n}ski, Filip and Wr{\\'o}blewska, Anna and Lipi{\\'n}ski, Dawid and Kaliska, Agnieszka and Rosalska, Paulina and Topolski, Bartosz and Biecek, Przemys{\\l}aw},
  journal={arXiv preprint arXiv:2105.05796},
  year={2021}
}

Citation

If you use the original Kleister benchmarks, cite the paper above. If you use this CTERA Hugging Face export, also cite this dataset card and specify the revision you used:

@dataset{orgrctera_kleister_charity_ie,
  title={Kleister Charity Information Extraction (CTERA format)},
  author={{CTERA AI}},
  year={2025},
  url={https://huggingface.co/datasets/orgrctera/kleister_charity_information_extraction}
}

(Adjust year and author string to match your internal release policy.)

Limitations

• This card describes the packaged columns in this repository; raw PDFs may be distributed separately under the original Kleister terms—consult applicaai/kleister-charity for the authoritative license and file layout.
• Metric computation should follow your agreed protocol when expected_output is empty on a split.