teplitsa-soc-tech/factbutcher-benchmark

English · Русский

FactButcher Russian Fact-Checking Dataset

423 claims in Russian, each with a reference verdict: true, false, or partly true. Use it to test how well a model, a prompt, or a service checks facts: give it the same claims and compare its answers with the reference ones.

We built the dataset to choose a model and settings for FactButcher.com, an AI fact-checking service. What that comparison showed is described below.

Examples

Claim	Verdict	Origin
Слоны боятся мышей (Elephants are afraid of mice)	False	Request from a FactButcher user
На западе Китая в чай кладут соль (In western China, people put salt in tea)	True	Request from a FactButcher user
Микеланджело говорил, что берет глыбу мрамора и отсекает от нее все лишнее. (Michelangelo said he takes a block of marble and cuts away everything unnecessary.)	False	Provereno.Media

All claims are in the CSV table. You can open it in a browser or in Excel, Google Sheets, or LibreOffice. To check a claim yourself before seeing the answer, hide the answer columns: gold_verdict and acceptable_verdicts.

Where the claims come from and who assigned the verdicts

Source	Claims	Who assigned the reference verdict
Requests to the FactButcher Telegram bot	274	For 100 claims, a person assigned the verdict based on the language models' checks. For the other 174, the models mostly assigned the verdict: two checked each claim independently with web search, and when they disagreed, a third checker resolved it. A person reviewed the final wording and verdicts
Provereno.Media articles	149	Provereno.Media's professional fact-checkers. We mapped their verdict to the dataset's scale and double-checked the mapping

Claims from user requests are not verbatim: we took the checkable statement out of the request and, when needed, added context from the same request, such as a date or place. The dataset contains no original messages and no data about users. Every Provereno.Media claim links to the published fact-check.

Collection and labeling are described in detail in METHODOLOGY.md.

How the verdicts work

• True (TRUE): the claim is supported.
• False (FALSE): the claim is contradicted.
• Partly true (MIXED): important parts of the claim differ in truthfulness, or reliable sources do not give one clear answer.

For some claims, two neighboring verdicts can both be honestly defended, for example "false" and "partly true". The main verdict is then in gold_verdict, and all accepted ones are in acceptable_verdicts. "Not enough evidence" (INSUFFICIENT_EVIDENCE) can be an accepted answer but is never the main verdict.

What our benchmark showed

In July 2026 we ran models from OpenAI, Anthropic, Google, and Perplexity through the dataset with different settings. Each model received one claim per request, searched the web itself, and returned a verdict, a short explanation, and links.

Selection had three stages: 15 configurations on 10 claims, then 9 on 100, and 5 finalists on all 423. A configuration here means a model together with its settings: how deeply it reasons and how much it searches the web.

Finalist	Share of accepted answers
OpenAI GPT-5.4-mini, tuned settings	81.6%
Perplexity Sonar Reasoning Pro	80.6%
Anthropic Claude Haiku 4.5	79.9%
OpenAI GPT-5.6 Luna	78.7%
OpenAI GPT-5.4-mini, default settings (how FactButcher ran during the test)	72.8%

• The difference between the top three is not statistically significant; there is no clear winner.
• Settings mattered more than the choice of model: the same GPT-5.4-mini with tuned settings answered correctly almost 9 percentage points more often than with default ones.
• Google Gemini models were dropped: they barely searched the web and cited made-up sources.

How we scored, which models were dropped and why, and what one check costs are in RESULTS.md.

Run the dataset yourself

1. Load the data. Download the CSV or JSONL file, or load it with the Hugging Face datasets library:

   from datasets import load_dataset
   
   rows = load_dataset("teplitsa-soc-tech/factbutcher-benchmark", split="test")
   

2. Send each claim from the claim field to the model or service you are testing. Save the answer together with claim_id.

3. Map the answers to four labels: TRUE, FALSE, MIXED, INSUFFICIENT_EVIDENCE.

4. Score the result in two ways:

   • accepted answers: the share of rows where the answer is in acceptable_verdicts (our numbers above are computed this way);
   • strict: the share of rows where the answer equals gold_verdict.

   Count rows the system did not answer as errors.

To make your result comparable with ours, run all 423 rows under the same conditions and report the model, prompt, and search settings. Also report whether provereno.media was reachable: a system with web search may find the published fact-check there. In our final numbers, Provereno.Media claims for which the models visited that site were re-run with the site blocked.

The dataset has one split, test. Do not use these rows to train or tune a system and then publish its result as an independent evaluation.

Fields

Field	Contents
claim_id	Unique claim identifier
claim	Claim to check
gold_verdict	Main reference verdict
acceptable_verdicts	All verdicts counted as correct. In the CSV they are separated by a vertical bar; in the JSONL this is a list
benchmark_component	Source: factbutcher_human_benchmark (user requests) or provereno_media
reference_date	Date the verdict applies to, when the claim depends on time
source_name	Name of the row's source
source_url	Provereno.Media article link; empty for user requests
source_license	License of the source material, when applicable
source_license_url	Link to that license

The full machine-readable field specification is in metadata/schema.json.

Limitations

• The collection is small and Russian-only. Topics are whatever FactButcher users brought and Provereno.Media covered; they were not balanced.
• Some claims depend on time; reference_date shows the date the verdict applies to.
• A reference verdict can still be contestable. acceptable_verdicts captures only part of that ambiguity.
• Reference verdicts for user requests rely heavily on language models. Mistakes shared by those models and a tested model may go unnoticed.
• Claims from user requests have no written fact-check and no complete list of sources.

Other limitations are listed in METHODOLOGY.md.

License and citation

The dataset is available under Creative Commons Attribution 4.0 International. Provereno.Media rows link to the original articles; see NOTICE.md for details. Citation metadata is in CITATION.cff.