SWARM is released for non-commercial research use only. The documents originate from public search engine results and may be subject to the source engines' terms of service and to third-party copyright. By requesting access you agree to use the data solely for research, not to redistribute it, and to honour takedown requests. Full terms are in the LICENSE file in this repository.

SWARM

SWARM (Search-Web documents Annotated for Russian propaganda, Multilingual) is a multilingual, human-annotated dataset for detecting support for Russian propaganda narratives in search engine results.

It contains 2,129 documents returned by four major search engines in response to queries about twenty recurring pro-Kremlin narratives, across nine languages (Arabic, English, German, Hindi, Polish, Portuguese, Russian, Spanish, Ukrainian). Each document was annotated by trained human coders for whether it supports a given narrative, yielding narrative-specific labels rather than broad propaganda-presence judgments.

Content warning. This dataset contains propaganda narratives, some with dehumanising elements, reproduced only to document and study the phenomenon.

Task

Given a document returned by a search engine and a narrative phrased as a single yes/no question, decide whether the document supports the narrative (answer to the question is yes). A document counts as supportive when it openly endorses the narrative (presenting it as fact or amplifying it uncritically), not merely when it mentions the topic. The negative class groups outright denial together with neutral or unrelated coverage, so the operative distinction is support versus everything else.

This is related to stance detection but harder: the target is a diffuse narrative rather than a sharply bounded proposition, and a document can support a narrative without asserting any single proposition. It also differs from fact verification, whose target is the veracity of a claim: a document can advance a propaganda narrative while making claims that are literally true, unfalsifiable, or value-laden.

Data fields

Single split (test), one row per document:

field	type	description
`id`	int64	Unique document identifier
`lang`	string	Query language (`ar`, `de`, `en`, `es`, `hi`, `pl`, `pt`, `ru`, `uk`)
`narrative_question`	string	The pro-Kremlin narrative operationalised as a yes/no question (English)
`label`	int64	1 if the document supports the narrative, 0 otherwise
`text_en`	string	Document text in English (original for `en`, machine-translated otherwise), truncated to 12,000 characters. Null for 15 documents (see Known limitations)
`text_native`	string	Document text as retrieved, truncated to 12,000 characters. Never null
`content_hash`	string	Hash of the document content (for deduplication / joining)

Both text_en and text_native are provided so the benchmark can be run either on English-normalised input or on native-language input. The labels and questions are identical across the two.

lang is the language of the search query, not of the document. Sampling is stratified by the language in which the search engine was queried, and engines routinely return results in other languages: a Hindi-language query, for instance, frequently returns English pages. Do not read lang as a document-language label, and do not use it as ground truth for language identification.

Statistics

2,129 documents, 384 positive (18.0%)
25 distinct narrative-question phrasings covering 20 pro-Kremlin narratives, organised around five actor groups (the United States, "the West", the European Union, Russia, Ukraine)

lang	n	positive	positive %
ar	258	46	17.8
de	244	52	21.3
en	217	44	20.3
es	227	34	15.0
hi	210	17	8.1
pl	244	24	9.8
pt	239	41	17.2
ru	244	89	36.5
uk	246	37	15.0

Usage

from datasets import load_dataset

ds = load_dataset("manueltonneau/SWARM", split="test")
ex = ds[0]
print(ex["narrative_question"])   # e.g. "do sanctions against russia primarily harm ..."
print(ex["label"])                # 1 = supports the narrative
print(ex["text_en"][:500])

Report metrics per language rather than only in aggregate: search results are a snapshot of specific regions and a time window, and both models and coders carry language-dependent behaviour. Positive-class F1 on the minority (support) class is the headline metric.

How it was built

Twenty narratives drawn from recurring themes in Russian state media were each turned into two search queries (one propaganda-framed, one neutral) and one yes/no question, translated into nine languages by native speakers. Search engines were audited with these queries; returned results were sampled, the linked pages/PDFs scraped, and each document annotated by trained coders for narrative support (two coders, or three after adjudication). Document text is truncated to 12,000 characters, matching the truncation used in the paper's benchmarks.

Known limitations

The dataset is released with its residual extraction and annotation noise intact rather than silently filtered, so that the item count matches the benchmark reported in the paper. Users running their own evaluation may wish to exclude the affected records.

Failed translations (15 documents). The translation pass that produced text_en occasionally returned a meta-response (a refusal, or a request to shorten the input) instead of a translation. Those strings are not document content, so text_en is set to null for the 15 affected records. Their text_native is intact and their labels are unaffected, so they remain usable for native-language evaluation. All 15 are label 0. Filter with ds.filter(lambda x: x["text_en"] is not None) when evaluating on English input.
Extraction failures (roughly 15 documents). A small number of records are bot-protection interstitials, "access denied" pages, or consent walls rather than article text, having passed the extraction filter because they contain a well-formed body of prose. They were annotated as presented. Most carry label 0, but three carry label 1.
Label conflicts (2 pairs). Four records form two pairs with identical text and identical narrative_question but opposite labels. All four are bot-protection pages, so the coders were in effect annotating boilerplate rather than content. They are retained for transparency.
Near-duplicate text. Twelve groups of records (25 rows total) share a content_hash, mostly because the same page was returned by more than one engine or region. Deduplicate on content_hash if independence across items matters for your analysis.
Selection effect on the positive class. Extraction does not succeed uniformly across sources. Russian state-linked domains are over-represented among the failures, and because such domains are more likely to carry propaganda-supporting content, the retained set may slightly under-count the positive class.

Intended use and access

SWARM is intended for non-commercial research on propaganda detection, multilingual NLP, and search-engine auditing. Because the documents come from public search results and may be subject to source engines' terms of service and third-party copyright, the dataset is distributed under gated, research-use-only access with a request process and a takedown procedure. It is not intended for training or deploying systems that generate, amplify, or target propaganda.

Citation

@misc{tonneau2026swarm,
  title  = {SWARM: A Multilingual Human-Annotated Dataset for Russian Propaganda Detection in Search Engine Results},
  author = {Tonneau, Manuel and Dubey, Abhinav and Shaikh, Farhan and Vitulano, Ilaria and Stolze, Martha and Dedeoglu, Hale and Riechert, Clara and Kuka, Ella and Sydorova, Maryna and Makhortykh, Mykola and Kuznetsova, Elizaveta},
  year   = {2026}
}

Downloads last month: 11

Total file size:

17.3 MB