This dataset contains records of real, named individuals and organisations taken from public sanctions, watchlist and politically-exposed-person sources. Access is granted automatically once you accept the licence and the terms of use below.

Sanctions Entity Resolution Benchmark

A benchmark for evaluating entity resolution (entity matching / record linkage) on sanctions and OSINT data. Each record is a pair of entities (persons or organizations) with a binary label indicating whether they refer to the same real-world entity.

This dataset is a snapshot of the OpenSanctions Pairs corpus (https://www.opensanctions.org/docs/opensource/pairs/), redistributed here in unmodified form to provide a stable, citable artefact for peer review and reproducibility. All credit for the underlying data belongs to OpenSanctions and the upstream curators whose source lists are aggregated by their cross-referencing pipeline. We claim no ownership of the data; this repository exists only to pin a specific snapshot used in our experiments.

Paper: OpenSanctions Pairs: A Large-Scale Dataset for Pairwise Entity Matching, NeurIPS 2026 Evaluations and Datasets Track.

Authors: Chandler Smith (University of Oxford), Magnus Sesodia (University of Oxford), Friedrich Lindenberg (OpenSanctions), Christian Schroeder de Witt (University of Oxford). Chandler Smith and Magnus Sesodia contributed equally. This repository is maintained by the authors; the Hugging Face account name sanctions-er-anon dates from double-blind review and is kept so that the URLs in the paper stay valid.

Code

Reference implementations, baselines, and evaluation scripts are available at:

https://github.com/chansmi/OSINT_entity_resolution

Dataset Summary

Pairs: 755,540
Positive (same entity): 76.9%
Negative (different entity): 23.1%
Entity types: Person, Organization
Languages: multilingual names; English metadata
Source: OpenSanctions Pairs corpus, snapshot dated 2025-12-09 (https://www.opensanctions.org/docs/opensource/pairs/)

Files

File	Description	Size
`pairs.json.gz`	Full benchmark, JSON Lines (gzipped)	~390 MB
`sample_1000.json`	1,000-pair representative sample (see below)	~3 MB
`clusters/clusters.jsonl.gz`	Canonical cluster of every entity (see below)	~23 MB

Cluster structure

clusters/clusters.jsonl.gz (config clusters) has one row per entity: entity_id, cluster_id (the smallest entity id in its cluster) and cluster_size. A cluster is a connected component of the graph whose edges are the positive pairs, so it groups every record that refers to the same real-world entity: 1,002,093 entities in 459,763 clusters, of which 35,341 contain more than two entities (up to 81). Use it for n-ary evaluation such as clustering and transitivity. It is produced by eval/export_clusters.py in the code repository.

1,000-pair representative sample

sample_1000.json is a label-stratified random subsample of the full 755,540-pair corpus, provided for quick inspection, smoke tests, and compliance review without downloading the full archive. It was generated by scripts/create_proper_sample.py with seed 42, as follows:

Stratification. The corpus is split by the analyst judgement field into positives and negatives. Quotas are set to match the corpus class balance: 76.9% positive (769 pairs) and 23.1% negative (231 pairs), giving 1,000 pairs total.
Random sampling. Within each class, pairs are drawn uniformly at random without replacement using Python's random.sample under a fixed seed (42).
Shuffling. The two class samples are concatenated and globally shuffled with the same seed so positives and negatives are interleaved rather than blocked.
Validation. The script verifies that (a) the realised positive ratio is within 0.02 of the 76.9% target, (b) the number of judgement-label runs is at least 70% of the expected value for an i.i.d. shuffle, and (c) every non-overlapping 100-pair window contains both classes. The released sample passes all three checks.
Provenance. The output JSON carries a metadata block recording the seed, target and realised positive ratios, source filename (pairs-20251209.json.gz), and creation timestamp, so the sample can be regenerated bit-for-bit from the public corpus.

The sample preserves the corpus-level class balance and is drawn uniformly across the entire labeled pool, so it inherits the schema mix, source-heterogeneity, and script-coverage distributions of the full release without any further filtering or reweighting. Full reproduction code is at https://github.com/chansmi/OSINT_entity_resolution.

Record Schema

One JSON object per line. The example uses invented values from the synthetic preview (https://github.com/chansmi/OSINT_entity_resolution/blob/main/data/preview/synthetic_pairs.json); real records have the same structure:

{
  "left": {
    "id": "synthetic-0021",
    "caption": "Lomfen Teltelis",
    "schema": "Person",
    "properties": {
      "birthDate": ["1974-08-16"],
      "nationality": ["xb"]
    }
  },
  "right": {
    "id": "synthetic-0022",
    "caption": "Ломфен Телтелис",
    "schema": "Person",
    "properties": {
      "alias": ["LOMFEN TELTELIS"],
      "birthDate": ["1974-08-16"]
    }
  },
  "judgement": "positive"
}

left, right: entity records with stable IDs, a display caption, a schema (Person or Organization), and a properties dict of attribute to list-of-string-values
judgement: "positive" (same entity) or "negative" (different)

properties keys vary by entity but commonly include alias, birthDate, nationality, country, address, idNumber, registrationNumber, incorporationDate.

Loading

The repository defines two configurations: default (the full pairs.json.gz) and sample_1000 (the 1,000-pair sample, read from the pairs field of sample_1000.json).

from datasets import load_dataset

ds = load_dataset("sanctions-er-anon/opensanctions_pairs", split="train")
print(ds[0])

sample = load_dataset("sanctions-er-anon/opensanctions_pairs", "sample_1000", split="train")

Or stream without download:

ds = load_dataset("sanctions-er-anon/opensanctions_pairs", split="train", streaming=True)

Access and Versioning

Access is gated: downloading requires a Hugging Face account and accepting the licence and terms of use shown on this page. Approval is automatic.

The current release is v1, a fixed snapshot of the OpenSanctions pairs file dated 2025-12-09. A released version is never rewritten, so results that cite it stay reproducible. Pin it when loading:

ds = load_dataset("sanctions-er-anon/opensanctions_pairs", split="train", revision="v1")

No further releases are currently planned. Because v1 is a snapshot, a person or organisation delisted after 2025-12-09 still appears in it. Do not use this dataset to make or support decisions about any individual; use the live OpenSanctions data for any operational purpose. If a new version is released, it will reflect upstream delistings and removals, and the terms of use require moving to it and deleting superseded copies.

Personal Data and Legal Basis

The records name real people and organisations, including names, aliases, dates of birth, nationalities and positions, because matching identities is the task itself; removing them would remove the benchmark. They are drawn only from official sanctions lists, regulatory watchlists and politically-exposed-person sources that authorities publish so that the designated parties can be identified. The dataset contains no records about the general public and adds nothing the sources did not publish.

The data is compiled by OpenSanctions Datenbanken GmbH (Berlin), which processes it under the EU General Data Protection Regulation on the basis of legitimate interest (Art. 6(1)(f) GDPR); see https://www.opensanctions.org/docs/privacy/. Requests to correct or delete a record go to the governmental or international body that maintains the relevant list, and a person removed from a source list is removed from OpenSanctions at its next reconciliation. Contact points for each listing body: https://www.opensanctions.org/docs/delisting/

Intended Uses

Benchmarking entity resolution / matching systems
Evaluating LLM and fine-tuned model performance on real-world noisy multilingual entity records
Studying compliance, sanctions screening, and OSINT use cases

Out-of-Scope and Limitations

Entities are drawn from sanctions, watchlist, and PEP sources; distribution is not representative of general-purpose record linkage (commercial CRM, healthcare, etc.).
Labels reflect cross-referencing decisions made by upstream curators and may contain noise.
Names are romanised / transliterated inconsistently across sources; some pairs are intentionally challenging due to script and transliteration variation.
The benchmark should not be used as a basis for real-world sanctions enforcement or watchlist screening without human review.

Ethical Considerations

The underlying data covers individuals and organisations on public sanctions and watchlists. All records are already public. Users should not redistribute the data in contexts that imply guilt or wrongdoing, and should follow the source licence and OpenSanctions terms of use.

Licence

Released under CC-BY-NC-4.0, inheriting the licence of the upstream OpenSanctions data. The dataset is for non-commercial research and evaluation use only; commercial use requires a separate licence from OpenSanctions (https://www.opensanctions.org/licensing/).

Citation

If you use this dataset, please cite OpenSanctions as the source of the underlying data. A pointer to the specific snapshot:

OpenSanctions. Pairs: Cross-Referencing Training Data.
https://www.opensanctions.org/docs/opensource/pairs/
Snapshot dated 2025-12-09.

If you use the benchmark, please also cite the paper:

@inproceedings{smith2026opensanctionspairs,
  title     = {OpenSanctions Pairs: A Large-Scale Dataset for Pairwise Entity Matching},
  author    = {Smith, Chandler and Sesodia, Magnus and Lindenberg, Friedrich and Schroeder de Witt, Christian},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS) Evaluations and Datasets Track},
  year      = {2026}
}

Source and Credit

This dataset is a verbatim snapshot of the OpenSanctions Pairs corpus, taken on 2025-12-09:

Pairs documentation: https://www.opensanctions.org/docs/opensource/pairs/
OpenSanctions homepage: https://www.opensanctions.org/
Source code and pipeline: https://github.com/opensanctions/opensanctions
Licensing: https://www.opensanctions.org/licensing/

OpenSanctions is the originator and maintainer of the data. The pairs are produced by their cross-referencing pipeline, which aggregates public sanctions lists, PEP registers, and watchlists from numerous upstream authorities. Please cite OpenSanctions in any work that uses this dataset, and consult their site for the canonical, continuously updated version.

We host this snapshot solely to provide a stable, immutable reference for reproducible benchmarking, since the live OpenSanctions data changes daily.

Downloads last month: 319

Total file size:

437 MB