idealab-cs2/cfgen-claims

CFGen: true and false claims generated from news articles

The labels say whether a claim agrees with its source article, not whether it is true in the world. Every claim was written by an LLM that had been told to produce some statements that match a news article and some that contradict it. is_factual records which of the two the model was asked for. Nobody checked these labels claim by claim, and the claims are not filtered for quality: in a 200-claim human check, about two thirds were judged clear and self-contained enough to fact-check.

19,202 claims about 336 news articles (January–March and September–October 2025, five languages), written by 9 open-weight LLMs. Five of them generated claims in 2025 and are kept as they were. In October 2026 each was matched with a current open model of similar size (four models, one of them run with and without its reasoning mode), and those were run on every article with the same prompt. 200 of the claims carry binary human labels from two annotators, plus the labels of 7 LLM judges on the same task.

The dataset is part of CFGEN (Contemporary Fact Generation Engine), a project on building fact-checking benchmarks from fresh news automatically: because the articles postdate the training cut-off of many models, a model can only verify a claim about them by reading the source or searching for it.

Dataset summary

Claims	19,202 (6,610 from the 2025 generators, 12,592 from the 2026 generators)
Articles	336, in three article sets (one config each)
Languages	English, German, Spanish, French, Italian
Publishers	39 news outlets in the US, UK, Germany, France, Spain and Italy
Article dates	2025-01-05 → 2025-03-25 and 2025-09-05 → 2025-10-30
Generators	2025: Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct, Qwen2.5-32B-Instruct, QwQ-32B, Qwen3-8B · 2026: gemma-4-12B-it, Qwen3.5-9B, Qwen3.8-27B (thinking off and on), Qwen3.5-122B-A10B
Labels	is_factual for every claim (what the generator was asked to produce); human is_checkable labels for 200 claims
Human agreement	Cohen's κ = 0.740 between the two annotators (199 claims), 0.773 on the 99 labelled independently
Article text	not included (see Source articles)

Configs and splits

One config per article set, one split per generator. A split name is the generator's name, with _thinking added when the model's reasoning mode was switched on.

2025q4_us: 85 US articles, September–October 2025

Six US outlets (Fox News, CNBC, The Washington Times, Rolling Stone, The New Yorker, The Washington Free Beacon), collected with Fundus. The only set that keeps article URLs.

split	generator	status	claims	true	false	articles
qwen3_8b	Qwen/Qwen3-8B	original	510	255	255	85
gemma_4_12b_it	google/gemma-4-12B-it	regenerated	440	236	204	85
qwen3_5_122b_a10b	Qwen/Qwen3.5-122B-A10B	regenerated	504	255	249	85
qwen3_5_9b	Qwen/Qwen3.5-9B	regenerated	544	301	243	85
qwen3_8_27b	Qwen/Qwen3.8-27B	regenerated	484	283	201	85
qwen3_8_27b_thinking	Qwen/Qwen3.8-27B	regenerated	511	270	241	85

2025q1_multilingual: 101 articles in five languages, January–March 2025

30 outlets in Italy, the UK, France, Spain, Germany and the US. Each article was used twice, with two prompt variants (method, see How it was made), so every split here mixes both.

split	generator	status	claims	true	false	articles
llama_3_1_70b_instruct	meta-llama/Llama-3.1-70B-Instruct	original	1,174	591	583	101
llama_3_1_8b_instruct	meta-llama/Llama-3.1-8B-Instruct	original	1,452	716	736	101
qwen2_5_32b_instruct	Qwen/Qwen2.5-32B-Instruct	original	1,205	607	598	101
qwq_32b	Qwen/QwQ-32B	original	1,289	689	600	101
gemma_4_12b_it	google/gemma-4-12B-it	regenerated	1,048	546	502	101
qwen3_5_122b_a10b	Qwen/Qwen3.5-122B-A10B	regenerated	1,255	690	565	101
qwen3_5_9b	Qwen/Qwen3.5-9B	regenerated	1,259	672	587	101
qwen3_8_27b	Qwen/Qwen3.8-27B	regenerated	1,051	574	477	101
qwen3_8_27b_thinking	Qwen/Qwen3.8-27B	regenerated	1,190	611	579	101

2025q1_english: 150 English articles, January–March 2025

17 UK and US outlets.

split	generator	status	claims	true	false	articles
llama_3_1_8b_instruct	meta-llama/Llama-3.1-8B-Instruct	original	980	479	501	150
gemma_4_12b_it	google/gemma-4-12B-it	regenerated	752	403	349	150
qwen3_5_122b_a10b	Qwen/Qwen3.5-122B-A10B	regenerated	903	460	443	150
qwen3_5_9b	Qwen/Qwen3.5-9B	regenerated	937	517	420	150
qwen3_8_27b	Qwen/Qwen3.8-27B	regenerated	830	481	349	150
qwen3_8_27b_thinking	Qwen/Qwen3.8-27B	regenerated	884	454	430	150

human_annotations: 200 claims, two annotators, 7 LLM judges

200 claims drawn from the 2025q4_us / qwen3_8b split (balanced 100 true, 100 false). See Human annotations.

Fields

Claim configs (2025q4_us, 2025q1_multilingual, 2025q1_english), one row per claim:

field	type	notes
claim_id	str	12-character hash of generator, mode, method, article and claim text
claim	str	the claim, in the article's language
is_factual	bool	true if the generator was asked for a claim that matches the article, false if it was asked for one that contradicts it
generator	str	Hugging Face id of the model that wrote the claim
generator_status	str	original (generated in 2025) or regenerated (October 2026)
method	str	naive (article only) or examples (three fact-checked claims from ClaimsKG shown first as style examples)
thinking	bool / null	whether the model's reasoning mode was on; null for the 2025 runs, where it was not a setting
reasoning	str	the model's own planning notes, written in the <REASONING> block the prompt asks for. Shared by all claims from the same answer. For thinking-mode runs, the hidden thinking trace is removed and only this block is kept
article_key	str	12-character hash of the article text; identical across splits, so claims about the same article can be joined
article_url	str / null	article URL (2025q4_us only; the 2025q1 crawl did not keep URLs)
article_title	str	the article's headline
publisher	str	news outlet
published	str	publication timestamp as given by the outlet
country	str	outlet's country (us, uk, de, fr, es, it)
language	str	ISO 639-1 code of the article and claim
generated	str	year-month the claim was generated

human_annotations has claim_id, claim, is_factual, generator, the article fields above, and:

field	type	notes
split	str	calibration (first 100, annotators discussed disagreements) or test (last 100, labelled independently)
is_checkable_annotator_1, is_checkable_annotator_2	0 / 1 / null	the two human labels (one claim has no label from annotator 1)
is_checkable_llm_qwen3_32b_v1	0 / 1	Qwen3-32B with the codebook prompt only, all 200 claims
is_checkable_llm_<judge>	0 / 1 / null	seven judges with the codebook plus 10 labelled calibration examples; test split only, so null on calibration and where the reply could not be parsed

How it was made

1. Articles. News articles were collected with Fundus from US, UK, German, French, Spanish and Italian outlets and grouped into the three sets above. One article in the multilingual set had an empty body; its claims (made up from nothing) are dropped from every split.
2. Prompt. Each model got the article's text and publication date with this instruction (system prompt: "You are an expert fact generator..."): write 1–3 claims that are clearly true according to the article and 1–3 that are false and contradict it; every claim must stand on its own, with the who, where and when spelled out (falling back to the publication date, e.g. "As of 2025-02-11, ..."), be plausible, and be in the article's language. The model first reasons in a <REASONING> block, then lists claims under <FACTUAL_CLAIMS> and <NON_FACTUAL_CLAIMS>. The full prompt is in scripts/claim_generation.py of the project repository.
3. Two variants on the multilingual set: naive (the prompt above) and examples, which first shows three claims with TRUE/FALSE verdicts from ClaimsKG, a collection of claims checked by fact-checking sites, as examples of style. The other two sets used naive only.
4. Parsing. Claims are read from the two lists. Answers that do not follow the format yield no claims for that article.
5. Regeneration (October 2026). Each 2026 model answered all 437 prompts (336 naive + 101 examples) once, with vLLM, temperature 1.0, top-p 0.95, seed 0, up to 8,192 new tokens (24,576 for Qwen3.8-27B with thinking on). All 2,185 answers ended on their own (none hit the token limit) and all of them produced at least one claim of each kind. The 2025 claims were generated with the same prompt by the earlier project code (April, July and November 2025); their exact sampling settings are not part of this release.

Which models, and why they were replaced

The 2025 generators were state of the art when they ran but are a year or more old now. Each was kept, and paired with the newest open model of comparable size:

2025 generator (release)	2026 counterpart (release)	why this one
Llama-3.1-8B-Instruct (Jul 2024)	gemma-4-12B-it (2026)	Meta has released no newer small Llama; the closest current dense model from a non-Qwen lab
Qwen3-8B (Apr 2025)	Qwen3.5-9B (Feb 2026)	direct successor, same size
Qwen2.5-32B-Instruct (Sep 2024)	Qwen3.8-27B, thinking off (2026)	newest dense Qwen of that size
QwQ-32B (Mar 2025, reasoning model)	Qwen3.8-27B, thinking on	Qwen folded QwQ's step-by-step reasoning into the thinking mode of its hybrid models
Llama-3.1-70B-Instruct (Jul 2024)	Qwen3.5-122B-A10B (Feb 2026)	there is no current dense ~70B model; this mixture-of-experts model is the class that fits the same hardware (one node of 4 GPUs)

The 2026 models are run on all three article sets, not only the one their 2025 counterpart covered, so every set can be compared across the same five 2026 configurations.

Human annotations

Two annotators labelled 200 claims from the 2025q4_us / qwen3_8b split with one binary question: is this a high-quality, checkable claim? Yes (1) means the claim is self-contained and clear (who, what, where and when are all given, no pronouns or references that need the article) and verifiable (an objective statement that could be proven true or false, not an opinion, prediction or hypothetical). It does not ask whether the claim is true.

The first 100 claims (calibration) were labelled together, with disagreements discussed to settle the codebook; the last 100 (test) were labelled independently. Annotator 1 judged 134 of 199 claims checkable, annotator 2 132 of 200; both said yes to 121.

The same question was put to LLM judges, prompted with the codebook. The _v1 column is a first Qwen3-32B run with the codebook alone; the seven other judges also got 10 labelled examples from the calibration claims and were scored on the test claims.

pair	split	n	Cohen's κ	accuracy
annotator_1 vs annotator_2	calibration	100	0.708	0.870
annotator_1 vs annotator_2	test	99	0.773	0.899
annotator_1 vs annotator_2	all	199	0.740	0.884
Qwen/Qwen3-32B vs annotator_1	test	99	0.682	0.859
Qwen/Qwen3-32B vs annotator_2	test	100	0.648	0.840
Qwen/Qwen3-14B vs annotator_1	test	99	0.590	0.828
Qwen/Qwen3-14B vs annotator_2	test	100	0.654	0.850
Qwen/Qwen3-30B-A3B-Instruct-2507 vs annotator_1	test	99	0.649	0.838
Qwen/Qwen3-30B-A3B-Instruct-2507 vs annotator_2	test	100	0.572	0.800
Qwen/Qwen3-235B-A22B-Instruct-2507 vs annotator_1	test	99	0.593	0.808
Qwen/Qwen3-235B-A22B-Instruct-2507 vs annotator_2	test	100	0.643	0.830
NousResearch/Hermes-4.3-36B vs annotator_1	test	99	0.537	0.818
NousResearch/Hermes-4.3-36B vs annotator_2	test	100	0.564	0.820
google/gemma-3-27b-it vs annotator_1	test	99	0.360	0.778
google/gemma-3-27b-it vs annotator_2	test	100	0.310	0.740
Qwen/Qwen2.5-14B-Instruct vs annotator_1	test	99	0.484	0.788
Qwen/Qwen2.5-14B-Instruct vs annotator_2	test	100	0.397	0.740

On the test claims the best LLM judge, Qwen3-32B, reaches κ = 0.682 and 0.648 against the two annotators, who reach 0.773 against each other. A second run of Qwen3-32B with the identical setup reached κ = 0.604 against annotator 1 (instead of 0.682), so differences of that size between judges should not be read as a ranking. Each judge number comes from one run on 99–100 claims.

Source articles

Article text is not redistributed. The articles belong to their publishers, and their licences do not allow republishing. Each claim instead carries the headline, outlet, publication date and, for 2025q4_us, the URL, which is enough to find the article. Claims and reasoning notes sometimes quote the article: in a random sample of 800 regenerated rows (and, separately, 600 original rows) the longest verbatim stretch from the article was 138 characters in a claim and 193 in a reasoning note (202 and 141 in the original rows), i.e. a sentence or less.

Use

from datasets import load_dataset

claims = load_dataset("idealab-cs2/cfgen-claims", "2025q4_us", split="qwen3_5_9b")
human = load_dataset("idealab-cs2/cfgen-claims", "human_annotations", split="train")

# compare generators on the same articles
old = load_dataset("idealab-cs2/cfgen-claims", "2025q1_multilingual", split="qwq_32b").to_pandas()
new = load_dataset("idealab-cs2/cfgen-claims", "2025q1_multilingual", split="qwen3_8_27b_thinking").to_pandas()
both = old.merge(new, on=["article_key", "method"], suffixes=("_old", "_new"))

Limitations

• is_factual is the instruction, not a verified label. A "true" claim can misread the article, and a "false" claim can accidentally be true in the world (or only partly false). Nobody checked individual claims against their article, apart from the 200 human labels, which ask a different question (checkability).
• True relative to the article, not to the world. If an article was wrong, its "true" claims are wrong too.
• Not filtered for quality. About a third of the human-labelled claims were judged not checkable (vague, missing context, opinion). The human labels cover one generator on one article set; other generators were not human-checked.
• One answer per article and model, at temperature 1.0. A different seed gives different claims.
• Unbalanced coverage. The 2025 generators each cover one article set; the 2026 ones cover all three. The multilingual set has 101 articles (31 English, 19 Spanish, 17 each French, German and Italian), so per-language numbers rest on few articles.
• No URLs for the 2025q1 sets, which the crawl did not keep; headline, outlet and date are given instead.
• Agreement numbers are small-sample. κ values come from 99–100 claims per split and one run per judge; see the Qwen3-32B rerun above.

License

The claims, reasoning notes and labels are released under CC BY 4.0. They were generated by models with their own licences: claims from Llama 3.1 models are subject to the Llama 3.1 Community License (Built with Llama); for the others see each model's card on Hugging Face. The source articles remain the property of their publishers and are not included. ClaimsKG examples appear only in the prompts, not in the data.

Citation

@misc{cfgen_claims_2026,
  title        = {{CFGen}: True and False Claims Generated from News Articles},
  author       = {Lazzaroni, Ruggero Marino},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/idealab-cs2/cfgen-claims}}
}

Data curation, generation and evaluation by ruggsea, using compute resources of the Complex Social & Computational Systems (CS²) group, IDea_Lab, University of Graz, and the Austrian Scientific Computing (ASC) MUSICA cluster.