eoplumbum/v4_nuclear_power_articles

Dataset Card for Nuclear News V4 Dataset

Dataset Summary

The Nuclear News V4 Dataset is a multilingual dataset consisting of 33,104 unique news articles sourced from 12 online news platforms across the Visegrád Group (V4) countries — Poland, Czech Republic, Slovakia, and Hungary — published between 1998 and 2025. The goal of the dataset is to analyze media narratives surrounding nuclear energy in Central Europe. While the dataset does not contain human-annotated (golden) labels, each article and its headline have been automatically evaluated in zero-shot settings by three different Large Language Models (Gemini Flash 2.0 (Google DeepMind, 2024), ChatGPT-4o (OpenAI, 2024), and DeepSeek-V3 (DeepSeek AI, 2024)) with respect to:

• Overall sentiment toward nuclear energy
• Sentiment of the headline (pessimistic ↔ optimistic)
• Degree of sensationalism / alarmism in the headline

Prompt used for annotation is shown below:

Here is the headline: {header}  
Here is the body of the article: {article_body}  
1.	How does this article portray nuclear power?
Choose only one number from 0 to 5:  
5 = Very positive (emphasizes benefits, strong support)  
4 = Positive (favorable, but not strongly supportive)  
3 = Neutral (balanced, without clear positive or negative framing)  
2 = Negative (emphasizes risks, somewhat unfavorable)  
1 = Very negative (strongly critical, highlighting risks)  
0 = Article not related to nuclear power  

2.	Is the message of the headline pessimistic or optimistic? Choose only one number from 1 to 5
5 = Very optimistic (emphasizes positive outcomes, hopeful)  
4 = Somewhat optimistic (positive framing, but not overly strong)  
3 = Neutral (presents information without clear sentiment)  
2 = Somewhat pessimistic (hints at concerns or negative outcomes)  
1 = Very pessimistic (strongly negative framing, highlights risks or failure)  

3.	To what extent is the headline alarmist?  
Choose only one number from 1 to 5:  
5 = Not alarmist (calm, measured tone, minimal exaggeration)  
4 = Slightly alarmist (minor emphasis on urgency or risk)  
3 = Neutral (fact-based, without apparent emotional emphasis)  
2 = Moderately alarmist (not extreme, but highlights threats or dangers)  
1 = Strongly alarmist (exaggerated emphasis on danger or crisis)  
Respond with numbers only.

Supported Tasks

This dataset can be used to perform:

• 'sentiment analysis' using LLM's labels as a golden labels
• 'summarization' treating headlines as ground-truth summaries.
• 'topic modelling' since most articles were scraped based on specific nuclear-related keywords (presented below)

Languages

According to BCP-47, the articles are written in the native languages of the respective sources:

• pl-PL for Polish
• hu-HU for Hungarian
• cs-CZ for Czech
• sk-SK for Slovak

No translation was performed — all texts remain in their original language.

Dataset Structure

Data Instances

Each instance is constructed as follow:

{
  "id": "28501",
  "country": "Poland",
  "language": "pl",
  "date": "26-06-2017",
  "year": "2017",
  "source": "Fakt_pl",
  "url": "https://www.fakt.pl/pieniadze/seksualny-skandal-w-elektrowni-jadrowej/sktj1t0",
  "title": "Seksualny skandal w elektrowni jądrowej!",
  "article body": "Co ma wspólnego bikini z energią atomową? Nazwa skąpego kostiumu pochodzi od atolu na Pacyfiku, na którym Amerykanie przeprowadzali jedne z pierwszych prób bomby atomowej. Efekt wywierany przez kobiety w tym stroju miał być równie wybuchowy. Podobnymi skojarzeniami kierowało się najwidoczniej szefostwo czeskiej elektrowni atomowej w Temelinie, na której terenie zorganizowano sesję zdjęciową maturzystek w bikini. Dziewczyny konkurowały o staż w elektrowni. Tylko czy zgrabne ciało powinno stanowić kryterium przyjmowania pracownika do strategicznego zakładu? \"Maturantka roku\", to czeski konkurs piękności dla maturzystek. W tym roku jedna z jego odsłon miała miejsce w elektrowni jądrowej Temelin. \"10 absolwentek liceum pozowało w bikini* w chłodni kominowej elektrowni. Następnie na facebookowej stronie operatora - firmy CEZ, notowanej także na GPW w Warszawie - kandydatki były oceniane przez komentatorów. Zwyciężczyni miała wygrać 14-dniowy staż\" - czytamy na Bankier.pl. Pomysł, by na podstawie konkursu piękności wyłaniać stażystkę do pracy w elektrowni spotkał się z falą krytyki w mediach, a firma CEZ wydała nawet oświadczenie, w którym tłumaczyła, że to nie ona była organizatorką konkursu. Ostatecznie staż zaproponowano wszystkim 10 uczestniczkom sesji w bikini. Elektrownia jądrowa Temelin to jedna z dwóch czeskich elektrowni atomowej. Działa od 2002 r. Jej budowa trwała 21 lat, czyli więcej niż liczą sobie kandydatki na \"Maturantkę roku\". Zatrzymała złodzieja a miała na sobie tylko bikini Sukienki z Lidla na lato. Szokująca cena",
  "gemini_article_rating": 3,
  "gemini_headline_sentiment": 1,
  "gemini_headline_alarmistic": 1,
  "deepseek_article_rating": 3,
  "deepseek_headline_sentiment": 3,
  "deepseek_headline_alarmistic": 2,
  "gpt_article_rating": 0,
  "gpt_headline_sentiment": 1,
  "gpt_headline_alarmistic": 2
}

Data Fields

Each row in the dataset corresponds to a single news article. The dataset contains the following fields:

• id (string): Unique identifier for each article.
• country (string): Country of origin. One of four values: Poland, Hungary, Czech Republic, Slovakia.
• language (string): Language code following BCP-47. One of: pl, hu, cs, sk.
• date (datetime): Exact publication date of the article (DD-MM-RRRR).
• year (datetime): Year of publication in YYYY format (range: 1998 to 2025).
• source (string): Name of the news platform (12 unique values).
• url (string): Original URL of the article.
• title (string): Title of the article.
• article body (string): Full text of the article.
• *_article_rating (int): How the article portrays nuclear power.
• *_headline_sentiment (int): Sentiment of the headline.
• *_headline_alarmistic (int): How alarmist the headline is.

Note: Replace * with one of the model names: gemini, deepseek, or gpt.

Quick start

from datasets import load_dataset

# Load dataset
dataset = load_dataset("eoplumbum/v4_nuclear_power_articles")

# Print dataset examples
print("Train data:", dataset["train"][0])
print("Validation data:", dataset["validation"][0])
print("Test data:", dataset["test"][0])

Dataset creation

Curation Rationale

This dataset was created to enable large-scale, cross-national comparisons of nuclear energy discourse in Central and Eastern Europe. By including both supportive and critical perspectives from various media outlets, the dataset allows researchers to explore narratives, misinformation, framing, and public opinion regarding nuclear energy.

Source data

Articles were scraped from publicly available news websites in the V4 region. None of the websites were behind a paywall or restricted in any other way at the time of scraping. The selected media outlets represent a diversity of political leanings and journalistic styles. Since the data was fed directly into LLMs, no extensive preprocessing was performed, except for a URL deduplication check.

Keywords

Keywords used for article scraping

["atom", "nuclear failure", "reactor failure", "nuclear malfunction", "nuclear safety", "atomic power plant", "nuclear power plant", "atomic energy", "nuclear energy", "nuclear energy", "nuclear power sector", "history of nuclear accidents", "nuclear incident", "nuclear disaster", "disaster at a nuclear power plant", "nuclear oversight", "radiation protection", "radioactive waste",
"nuclear fuel", "nuclear policy", "radiation", "nuclear failure reports", "atomic reactor", "nuclear reactor", "research reactors", "development of nuclear energy", "radioactive contamination", "nuclear technologies", "energy transition", "nuclear accident", "nuclear accidents", "uranium resources", "nuclear fuel cycle", "energy transformation"]

Annotation process

After scraping articles based on keywords or archive dates, the assessment of their relevance to the nuclear topic was performed by a single model (Gemini Flash 2.0, Google DeepMind, 2024). This step filtered out unrelated articles before passing the remaining ones to other LLMs.

Normalization

The dataset has not been blanaced or normalized.

Personal and Sensitive Information

No personal or private user data is included.

Considerations for Using the Data

This dataset is intended to support informed public discourse, academic research, and media literacy regarding nuclear energy. It provides a foundation for analyzing cross-cultural differences in energy communication and political bias. It may also be used to analyze differences in opinions and biases between large language models (LLMs).

Discussion of Biases

As the data originates from various media outlets with differing political and editorial biases, coverage may be uneven across countries, time periods, or stances. Language bias is also present due to the varying availability of nuclear-related articles in each language. Since only one LLM in a zero-shot setting was used to filter for relevance, some articles may have been misclassified as unrelated and excluded.

Limitations

No manual checking (except for selected examples) has been performed. Temporal coverage may be uneven across countries, as some articles might have been excluded during filtering. Scraped articles may contain minor formatting errors.

Licensing Information

The dataset is distributed for research and educational purposes only. Rights to the original articles remain with the publishers