| --- |
| language: |
| - pl |
| pretty_name: Polish BookSummary |
| task_categories: |
| - text-classification |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - polish |
| - long-context |
| - document-classification |
| - factual-consistency |
| - literature |
| - translated |
| - synthetic |
| --- |
| |
| # Polish BookSummary |
|
|
| **Polish BookSummary** is a long-document factual-consistency classification dataset constructed from a publicly available collection of English-language book plot summaries. |
|
|
| Each example consists of a short **claim followed by a Polish plot summary**. The task is a **binary classification problem** in which the model must determine whether the claim is factually consistent with the information contained in the document. |
|
|
| Polish BookSummary is part of the **LongContext** benchmark introduced with [Polish ModernBERT](ARXIV_URL). |
|
|
| ## Dataset statistics |
|
|
| | Split | Examples | |
| |---|---:| |
| | Train | 4,460 | |
| | Validation | 620 | |
| | Test | 1,248 | |
| | **Total** | **6,328** | |
|
|
| The dataset contains predominantly long inputs: **87.9%** of test examples exceed 512 tokens, **38.7%** exceed 1,024 tokens, and **8.1%** exceed 2,048 tokens when tokenized with the Polish ModernBERT Base tokenizer. |
|
|
| ## Dataset construction |
|
|
| The dataset was constructed from a public collection of approximately **25.7K English-language plot summaries**. To obtain sufficiently long inputs, the **longest 25%** of the summaries were retained and translated into Polish using **GLM-4.6**. |
|
|
| For each summary, GLM-4.6 generated a **two- or three-sentence claim supported by the document**, with a preference for facts appearing in the second half of the summary. The claim was then prepended to the summary. |
|
|
| For half of the examples, the claim remains supported by the document. For the other half, the claim was modified by GLM-4.6 to make it plausible but factually inconsistent with the corresponding summary. The resulting dataset is therefore a balanced binary classification task requiring the model to verify information across a long document. |
|
|
| The task is evaluated using **accuracy**. |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite: |
|
|
| ```bibtex |
| [CITATION] |
| ``` |
