strebeyko/smigiel

ŚMIGIEL: Spotting Machine-Generated Text from LLMs for Polish

śmigiel ['ɕmiɡjɛl] m I, D. ~gla gw. «płaski szczebel w drabinach wozu» (Eng. flat rung in a ladder wagon) // SJP W. Doroszewskiego

Dataset Summary

Consisting of 64 538 human-written and machine-generated texts in Polish from various domains, ŚMIGIEL is a comprehensive resource for training and benchmarking Machine-Generated Text (MGT) detection systems focusing on Polish language. Here's the original ŚMIGIEL paper. 👨🏼‍🔬

The dataset was originally created to for needs of the Shared Task 1 at PolEval 2025 organized by The Linguistic Engineering (LE) Group, part of the Department of Artificial Intelligence at the Institute of Computer Science, Polish Academy of Sciences IPI PAN.

Try it

To load the dataset, the easiest is to use HuggingFace's own Datasets library

from datasets import load_dataset
ds = load_dataset("strebeyko/smigiel")

Data Overview

Fields

Column	Type	Values	Description
text	string	—	The text fragment to classify. Prefix-stripped, whitespace-normalized, and truncated.
model	string	human, bielik-sm, bielik-md, mistral-sm, mistral-md, plum, gemma, llama-sm, llama-lg	Source of the text — either human or the name of the LLM that produced it.
strategy	string	human, greedy, sampling, beam_search, contrastive, dbs, llama_plum_sampling	Decoding strategy used during generation. Set to human for human-written texts.
key	ClassLabel	0 = human, 1 = machine	Binary classification label.

For fuller model details, refer to models section. Strategy parameters are explained at strategies section.

Splits

The dataset is split between train, test_a and test_b.

Split	Size	Description
train	35,763	Human-written texts from 4 domains (Wikipedia, literature, social media, reviews) paired with completions from 7 models (3 small, 3 medium, 1 large).
test_A	10,343	Interleave of: ½ alpha test split + ⅓ beta (Llama 3.3 70B only) + ⅓ gamma (8 models, adds news & government domains). Released with labels during the competition.
test_B	18,432	Interleave of: ½ alpha test split + ⅔ beta + ⅔ gamma. Larger blind test set covering same domain/model mix as test_A. Labels released after the competition.

Curation Rationale

The dataset was originally compiled for Shared Task 1 at PolEval 2025 to

1. provide training data for SUPERVISED subtask;
2. provide two sets of testing data for benchmarking submissions;

Source Data

The original, human-written texts are coming from 12 distinct datasources. Collectively, they cover wide range of styles, lengths, and purposes of text-writing.

Internal name	Domain	Description	Source
wiki	wiki	Polish Wikipedia	chrisociepa/wikipedia-pl-20230401
plsc	literature	Polish scientific article abstracts	rafalposwiata/plsc
coursebooks	literature	Polish open coursebooks	rafalposwiata/open-coursebooks-pl
classics	literature	Polish classic literature corpus	dmitriilebedev/polish-corpus (Kaggle)
twitter	social	Polish tweets (TwitterEmo)	clarin-pl/twitteremo
wykop	social	Polish social media posts (BAN-PL, non-offensive only)	ZILiAT-NASK/BAN-PL
polemo_hotels	reviews	Hotel reviews (PolEmo 2.0)	clarin-pl/polemo2-official
polemo_medicine	reviews	Medical reviews (PolEmo 2.0)	clarin-pl/polemo2-official
polemo_products	reviews	Product reviews (PolEmo 2.0)	clarin-pl/polemo2-official
polemo_courses	reviews	Course reviews (PolEmo 2.0)	clarin-pl/polemo2-official
allegro	reviews	Allegro marketplace reviews	PL-MTEB/allegro-reviews
filmweb	reviews	Polish movie reviews (FilmwebPlus)	narolski/filmwebplus
pmrd	reviews	Polish Movie Reviews Dataset	kamilsan/polish-movie-reviews-dataset
wikinews	news	Polish Wikinews articles (custom scrape)	pl.wikinews.org
gov	government	Polish parliamentary debates — Sejm + Senat (ParlaMint 5.0)	ParlaMint-PL, CLARIN.SI

Domains

To help to balance the detaset in terms of linguistic features, the original sources were grouped into 6 genres or "domains". This balancing is two-fold - it's inward, as the origin of texts within a domain is balanced how much was possible, and outward, as the main part of the data, (train and alpha test subgroup) consists of equal share of the 4 domains. The two remaining ones, news articles and parlimentary hearings, were introduced as part of robust training subset.

Models

To provide for versitile MGT, we used models coming from different families, and of varying sizes.

Moniker	Size	Full name	HuggingFace
llama-sm	small	Llama 3.1 8B Instruct	meta-llama/Llama-3.1-8B-Instruct
bielik-sm	small	Bielik 7B Instruct v0.1	speakleash/Bielik-7B-Instruct-v0.1
mistral-sm	small	Mistral 7B Instruct v0.3	mistralai/Mistral-7B-Instruct-v0.3
bielik-md	medium	Bielik 11B v2.3 Instruct	speakleash/Bielik-11B-v2.3-Instruct
mistral-md	medium	Mistral Nemo Instruct 2407	mistralai/Mistral-Nemo-Instruct-2407
plum	medium	PLLuM 12B nc chat	CYFRAGOVPL/PLLuM-12B-nc-chat
gemma	large	Gemma 3 27B Instruct	google/gemma-3-27b-it
llama-lg	large	Llama 3.3 70B Instruct	meta-llama/Llama-3.3-70B-Instruct

Strategies

We foster versitality of (generated) data by applying different decoding strategies. These strategies condition how "next token candidates" or strings of thereof are ultimately selected by the model. Below we provide their rundown, together with how they translate into parametrs of model inference's call.

Strategy	Full name	Parameters	Reference
greedy	Greedy decoding	do_sample=False	HF docs
sampling	Multinomial sampling	do_sample=True, num_beams=1	HF docs
beam_search	Beam search	num_beams=2	HF docs
contrastive	Contrastive search	penalty_alpha=0.6, top_k=4	Su et al., 2022
dbs	Diverse beam search	num_beams=6, num_beam_groups=3, diversity_penalty=1.0	Vijayakumar et al., 2018
llama_plum_sampling	Temperature sampling	do_sample=True, temperature=0.6, top_p=0.9	— (custom config for llama-lg and plum)

Composition

Test A and Test B are both composed out of thee distinct subsets of data - alpha, beta, and gamma. The two tests differ in proportions of sampled data.

Subset	Data	Models	Examples	Share in test_a	Share in test_b
alpha	old (4 domains)	all 7 base models + human	4 505	½	½
beta	old (4 domains)	llama-lg only + human	4 356	⅓	⅔
gamma	new (news + gov)	all 8 models + human	19 914	⅓	⅔

• alpha is simply the test split of the base postprocessing run
• beta is a result of processing human texts from the dev split with a single model (llama-lg, Llama 3.3 70B)
• gamma uses all the available models to process data from unseen datasets (the data was not published before)

The subsets contribute to the two resulting test groups through a round-robin assignment method. While the contribution of alpha is equal in both, the larger Test B gets more content from beta and gamma.

Uses

ŚMIGIEL dataset is rich and balanced, making it a great fit for training supervised-learning systems in detection of GenAI text. The two tests subsets are robust and include data from domains and by models not seen in training, making it a great testing ground for all MGT-detecting architectures.

Direct Use

The dataset is useful whenever a use case:

• requires a big body of texts in Polish from various domains
• is concerned with comparing generations coming from different LLM families, accross varying decoding strategies
• needs a balanced representation of human-written and machine-generated texts to train and benchmark detection systems

Out-of-Scope Use

ŚMIGIEL is not for you, if:

• you seek to train and validate systems that define the task not as classification, but as a text-boundary problem. This definition often requires more nuanceced solutions and more granular data (see Named-entity Recognition).
• you seek using the dataset to enhance systems that pose as humans with intention to deceive. This is considered a malicious use of the dataset.

Bias, Risks, and Limitations

1. One of important limitations of all the datasets for training MGT detection systems, is that the data gets stale fast, due to being incorporated into LLM training input.
2. The authors realise that although the dataset is intended to aid machine-generated detection systems, it could be potenially used to advance systems meant to deceive humans. The authors consider it a misuse of the dataset.

Recommendations

Both limitations (staleness and risk of misuse) stem directly from the open-source nature of ŚMIGIEL and cannot be effectively mitigated.

FAQ

Who are the source data producers?

The source data comes from open-sourced corpora authors as well as researchers at IPI PAN to whom we are deeply indepted.

Personal and Sensitive Information

The usernames in texts coming from social media have been anonimized. To our best knowledge, the data does not contain any other personal and / or sensitive information.

Data Collection and Processing

The source data is processed to produce prompts to guide machine generation. The generations then further filtrated and sampled from. For details on the processes please refer to our repo.

Citation

@article{
  title={Śmigiel Dataset: Laying Foundations for Investigating Machine-Generated Text Detection in Polish},
  author={Jakub Strebeyko, Alina Wróblewska, Piotr Przybyła},
  journal={LREC 2025},
  year={2025},
  doi={10.63317/3p7ghe9pfm8v}
}

• Designed by: Piotr Przybyła, PhD and Alina Wróblewska, PhD
• Compiled and Curated by: JJ Strebeyko
• Language(s) (NLP): Polish
• License: Creative Commons Attribution 4.0
• Repository: https://github.com/JStrebeyko/code-of-smigiel
• Paper: PolEval 2025 Task 1 Śmigiel: Spotting Machine-Generated Text from LLMs for Polish
• Resources: Zenodo

Acknowledgements

Dataset creation was the Ramón y Cajal grant RYC2024-050327-I, funded by the Spanish State Research Agency (MI-CIU/AEI/10.13039/501100011033) and by the European Social Fund Plus (ESF+) of the European Union. We also gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Centers: ACK Cyfronet AGH) for providing computer facilities and support within computational grant no. PLG/2025/018019.