Polish Media Articles with Subjectivity Scores from Four LLM Judges

Polish articles from various media outlets, judged for subjectivity using four large language models. The dataset includes individual model reviews with reasoning and scores, as well as an aggregated average. It was created by Jakub Legutko and last updated on February 23, 2026.

Use Cases

• Benchmarking LLM performance on Polish text subjectivity tasks based on the four provided model outputs.
• Analyzing patterns of media bias across Polish outlets based on the aggregated LLM judgments.
• Training or fine-tuning models for subjectivity detection using the labeled article scores.
• Comparing reasoning patterns between specialized and general-purpose LLMs on a non-English task.

Strengths

• Includes judgments from four distinct LLMs (Bielik-11B, Llama-3-8B, Mistral-7B, DeepSeek-R1) allowing for comparative analysis.
• Provides both raw model outputs with reasoning and an aggregated average score for each article.
• Dataset is licensed as CC0-1.0, permitting unrestricted use.
• Includes a script (regenerator.py) to recreate the article text used for analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The description does not specify the number of articles, topics, or time period covered.

Provenance

Source

Jakub Legutko, via Dataverse.

Collection Method

Articles gathered from various Polish media outlets and judged for subjectivity using LLMs.

Freshness

Last updated 2026-02-23 17:45:29; freshness should be verified.

Geography

Poland