SumeCzech: Czech News Summarization Dataset

Updated 8 July 2026

SumeCzech is a comprehensive Czech news summarization dataset comprising roughly 1,000,000 article–summary pairs that support both headline and abstract generation.
It is meticulously preprocessed with language filtering, duplicate removal, and field-length checks to ensure high-quality, journalistic texts.
The dataset underpins state-of-the-art abstractive summarization research with models like mT5-SC and M7B-SC, and it facilitates transfer to other summarization domains.

Searching arXiv for the cited SumeCzech-related papers to ground the article in current sources. arXiv search query: (Tran et al., 14 Aug 2025) SumeCzech Czech summarization SumeCzech is a large-scale Czech news-based summarization dataset designed for supervised training and evaluation of Czech summarization systems. It was created at the Institute of Formal and Applied Linguistics, Charles University, and is described as a notable exception to the scarcity of Czech-specific resources, with roughly 1,000,000 article–summary pairs drawn from major Czech online news portals. The dataset supports both headline generation and multi-sentence abstract generation, although recent work has primarily used it for abstractive article-to-abstract summarization and has treated it as the main benchmark for modern Czech summarization (Tran et al., 14 Aug 2025).

1. Definition, scope, and institutional origin

SumeCzech is a modern Czech summarization corpus centered on journalistic text. Its primary purpose is supervised training and evaluation for Czech news summarization, and its domain is general-domain news articles from major Czech online news portals. In the more recent literature, it is characterized as “the most comprehensive dataset available for modern Czech summarization,” and as a foundational benchmark for contemporary Czech text summarization (Tran et al., 24 Nov 2025).

The dataset contains approximately 1,000,000 Czech news articles. The sources identified in the recent summarization studies are České Noviny, Deník, iDNES, Lidovky, and Novinky.cz. Each record is stored in JSONLines format and contains the fields url, headline, abstract, text, subdomain, section, and date (Tran et al., 14 Aug 2025).

SumeCzech supports two summarization targets. The first is headline generation, which corresponds to very short summaries. The second is multi-sentence abstract generation, in which the target is a short journalistic abstract. The recent large-language-model studies focus on the second setting, using the text field as input and the abstract field as output (Tran et al., 14 Aug 2025).

This positioning matters because SumeCzech is not merely a generic corpus of short news texts. It is explicitly structured as a paired summarization dataset with journalist-authored targets, and the resulting supervision encodes modern Czech journalistic compression practices. A plausible implication is that the dataset functions not only as a benchmark but also as a repository of stylistic conventions for Czech news abstraction.

2. Corpus structure, preprocessing, and linguistic profile

The preprocessing pipeline summarized in the recent work includes language recognition to filter out non-Czech content, duplicate removal for exact and near duplicates, and filtering of entries with empty headlines, abstracts, or texts, as well as entries with excessively short fields (Tran et al., 14 Aug 2025). This yields what the paper characterizes as a large-scale, relatively clean modern Czech news summarization corpus.

The standard split is approximately 86.5% training, 4.5% development, and 4.5% test. The average full-text length is about 409 words, while the average abstract length is about 38 words, so the abstract is roughly one tenth of the source length (Tran et al., 24 Nov 2025). The summaries are therefore short, compressive, and multi-sentence rather than headline-like in the experimental settings emphasized by recent work.

SumeCzech is linguistically defined by modern standard Czech and by mainstream journalistic style. This includes short, fact-focused abstracts summarizing who, what, when, where, and why. The dataset therefore differs sharply from historical Czech corpora such as Posel od Čerchova, which contain outdated vocabulary and in some cases German text (Tran et al., 14 Aug 2025).

Several properties of Czech shape the difficulty of the task. The recent literature frames Czech summarization as challenging because Czech is morphologically rich, with extensive inflection for case, number, gender, and verbal categories; because it permits relatively free word order; and because news texts are dense in named entities and Czech-specific domain vocabulary (Tran et al., 24 Nov 2025). These properties are intrinsic to SumeCzech and influence both modeling and evaluation.

3. Task formulation and evaluation regime

In current usage, SumeCzech is treated primarily as an abstractive summarization dataset. Its headlines and abstracts are authored by journalists rather than mechanically extracted from source documents, and they are frequently paraphrastic and compressive. Accordingly, recent systems formulate the task as article-to-abstract generation in Czech (Tran et al., 14 Aug 2025).

For sequence-to-sequence models, the input is the full article text and the output is the abstract. The standard training objective is token-level cross-entropy over the target summary sequence. In canonical form, recent work states the objective as

$\mathcal{L} = -\sum_{t=1}^{T} \log p_\theta(y_t \mid y_{<t}, x),$

where $x$ is the input article and $y_1,\dots,y_T$ are the target summary tokens (Tran et al., 14 Aug 2025).

Evaluation is based on ROUGERAW, specifically ROUGE $_{\text{raw}-1}$ , ROUGE $_{\text{raw}-2}$ , and ROUGE $_{\text{raw}-L}$ , each reported with precision, recall, and F1. ROUGERAW measures raw token-level overlaps without stemming, lemmatization, or other preprocessing. The papers emphasize that this makes the metric particularly strict for Czech, because different inflected forms do not count as matches (Tran et al., 24 Nov 2025).

This strictness has methodological significance. In morphologically rich languages, ROUGERAW rewards not only content selection but also morphologically accurate lexical realization. A plausible implication is that the metric partially conflates semantic adequacy with surface-form fidelity. The recent literature explicitly notes that semantically correct summaries may be undervalued if they differ morphologically or lexically from the journalist-written reference (Tran et al., 14 Aug 2025).

An earlier line of work on SumeCzech also used headline generation rather than abstract generation and introduced named-entity-focused evaluation. In that setting, the dataset was enriched with named entity annotations to produce SumeCzech-NER, and the paper proposed ROUGE $_{\text{NE}}$ , a metric measuring overlap of named entities between gold and generated summaries. That study argues that correct entity transfer remains difficult even when standard lexical metrics improve (Marek et al., 2021).

4. Modeling approaches and state-of-the-art results

Recent work evaluates large multilingual LLMs fine-tuned directly on SumeCzech. Two central systems are mT5-SC, an mT5-base model fine-tuned on SumeCzech, and M7B-SC, a Mistral 7B model fine-tuned on SumeCzech via QLoRA (Tran et al., 14 Aug 2025).

mT5-SC uses an encoder–decoder Transformer architecture pretrained on mC4 and tokenized with SentencePiece. The recent studies state that SentencePiece helps handle complex Czech morphology through subword segmentation. Fine-tuning uses AdamW with learning rate $0.001$ on a single NVIDIA A40 GPU (Tran et al., 24 Nov 2025).

M7B-SC uses a decoder-only Transformer with Grouped-Query Attention and Sliding Window Attention. Fine-tuning is performed with QLoRA, in which the 7B backbone is 4-bit quantized, low-rank adapters are attached, and only adapter parameters are updated while the original weights remain frozen (Tran et al., 14 Aug 2025).

The strongest previously reported Czech summarization baseline in the recent comparison is HT2A-S, an mBART-based model fine-tuned on SumeCzech. Older baselines include First, Random, TextRank, and Tensor2Tensor (Tran et al., 24 Nov 2025).

The reported SumeCzech test-set results establish a new state of the art for M7B-SC.

Method	ROUGE $_{\text{raw}-1}$ F1	ROUGE $_{\text{raw}-2}$ F1	ROUGE $x$ 0 F1
M7B-SC	21.2	5.7	15.5
mT5-SC	19.2	4.6	14.1
HT2A-S	18.2	4.6	13.5
TextRank	13.8	0.3	0.8
First	14.4	0.2	0.9
Random	12.7	0.1	0.8
Tensor2Tensor	11.3	0.1	0.8

The papers state that M7B-SC outperforms all baselines and achieves new state-of-the-art results on the dataset, while mT5-SC consistently obtains the second-best results (Tran et al., 14 Aug 2025). This demonstrates that direct Czech fine-tuning of multilingual LLMs is effective even under the strict ROUGERAW evaluation regime.

An earlier stage of SumeCzech research had already shown that named-entity-aware abstractive methods could surpass the original baselines. In the headline-generation setup, Seq2Seq and Seq2Seq–NER improved substantially over earlier systems, and Seq2Seq–NER showed slightly better out-of-domain performance, suggesting that entity-aware inputs can improve robustness (Marek et al., 2021).

5. Named entities, extractive baselines, and summary style

Named entities occupy a distinctive place in SumeCzech-based research because Czech news summaries are structurally organized around actors, institutions, places, and times. The named-entity study built SumeCzech-NER by annotating the dataset with entities using a SpaCy model trained on the Czech Named Entity Corpus 2.0, and released the resulting resource separately (Marek et al., 2021).

That study proposed Named Entity Density, an extractive method selecting the sentence with the highest ratio between the number of entity tokens and sentence length. The authors report that this approach reaches results close to the First-sentence baseline in the news domain and tends to select sentences that concisely identify to whom, when, where, and what happened (Marek et al., 2021).

The same paper introduced ROUGE $x$ 1, defined over named-entity tokens only. On the SumeCzech test set in the headline-generation setting, the reported ROUGE $x$ 2 F1 values are low: approximately 2.4 for First, 1.6 for Named Entity Density, 5.6 for Seq2Seq, and 5.0 for Seq2Seq–NER (Marek et al., 2021). The authors interpret these results as evidence that entity fidelity remains difficult even for systems that improve standard ROUGE scores.

This reveals an important aspect of SumeCzech. The dataset is not simply a benchmark for compressive lexical overlap; it is also a test of whether systems preserve the identity structure of the news event. In news summarization, an incorrect person, institution, or location can alter the proposition itself. A plausible implication is that entity-sensitive evaluation should remain central in future work, especially for factuality-oriented applications.

6. Transfer, contamination, and broader significance

SumeCzech has become a base domain for transfer to other forms of Czech summarization. The recent work on historical documents uses SumeCzech-trained models as starting points and explicitly models historical summarization as cross-domain or cross-temporal adaptation from modern Czech news style (Tran et al., 14 Aug 2025). The journalistic form of SumeCzech abstracts was even used to guide the construction of summaries for the historical Posel od Čerchova dataset, where prompts requested summaries “in a journalistic style” to mirror SumeCzech (Tran et al., 24 Nov 2025).

The dataset therefore has significance beyond its immediate benchmark role. It serves as a training bed for Czech summarization models, a stylistic template for other summarization resources, and an empirical basis for studying multilingual large-language-model adaptation to a morphologically rich medium-resource language.

At the same time, recent Czech benchmark work has raised concerns about contamination. BenCzechMark initially considered SumeCzech as a summarization task in a generation format evaluated with RougeRAW-2, but later removed it because contamination analysis against the BUT-Large Czech Collection found 97.07% contamination for SumeCzech abstracts (Fajcik et al., 2024). The benchmark authors argue that such contamination makes the dataset unsuitable for clean evaluation of web-scale Czech LLMs trained on broad crawled corpora.

This does not diminish SumeCzech’s value as a training resource or as a historical benchmark, but it does complicate its role in contemporary large-model evaluation. A plausible implication is that SumeCzech now occupies two distinct methodological positions: it remains indispensable for supervised Czech summarization research, yet it may no longer function as an uncontaminated test bed for models pretrained on large public Czech web corpora.

Future directions identified in the recent summarization literature include hybrid or ensemble architectures, instruction tuning, reinforcement learning for improved content selection and faithfulness, cross-temporal transfer from modern to historical Czech, and semantic evaluation metrics such as BERTScore that are less sensitive to morphological variation (Tran et al., 14 Aug 2025). These proposals collectively position SumeCzech as both a mature benchmark and a continuing point of departure for research on Czech summarization, morphology-sensitive evaluation, and domain transfer.