This dataset provides preprocessed Czech text data for training SLiM-CZ-V1 (Slavic Linguistic integrated Micro-model for Czechia), a small transformer-based language model designed for Czech text generation and language modeling tasks.

Dataset Details

Dataset Description

SLiM-CZ-V1 Czech Text Corpus contains tokenized Czech text sequences ready for training autoregressive language models. The dataset has been preprocessed with consistent cleaning, tokenization, and sequence creation to ensure high-quality training data for Czech language models.

Curated by: Filip Sedivy
Language(s) (NLP): Czech (cs)
License: MIT License

Dataset Sources

Repository: https://github.com/filipsedivy/SLiM-CZ-V1
Model Repository: https://huggingface.co/filipsedivy/SLiM-CZ-V1

Uses

Direct Use

This dataset is designed for:

Training small to medium-sized Czech language models (3M-125M parameters)
Autoregressive text generation in Czech
Language modeling research for Czech NLP
Fine-tuning pre-trained models for Czech-specific tasks
Educational purposes for understanding transformer-based language models

Recommended use: Training SLiM-CZ-V1 models (Tiny, Small, Medium, Large variants).

Out-of-Scope Use

This dataset should NOT be used for:

Production systems without human oversight
Medical, legal, or financial decision-making
Generating harmful or illegal content
Applications where factual accuracy is critical without verification
Training models for languages other than Czech

Dataset Structure

Data Format

The dataset is provided in JSON format as a list of token ID sequences:

[
  [15, 32, 45, 67, 89, 12, 34, 56, 78, 90, 23, 45, ...],  // Sequence 1
  [32, 45, 67, 89, 12, 34, 56, 78, 90, 23, 45, 67, ...],  // Sequence 2
  [45, 67, 89, 12, 34, 56, 78, 90, 23, 45, 67, 89, ...],  // Sequence 3
  ...
]

Each sequence is a list of integer token IDs with length seq_len + 1:

First seq_len tokens serve as input
Last seq_len tokens serve as labels (shifted by 1 position)

This structure enables autoregressive language modeling where the model predicts the next token.

Example

With seq_len=512:

Each sequence has 513 tokens
Input: tokens [0:512]
Target: tokens [1:513]
This creates a "next token prediction" task

Data Files

processed_data/
├── train.json       # Training sequences (list of lists)
├── val.json         # Validation sequences (list of lists)
├── test.json        # Test sequences (list of lists)
├── tokenizer.json   # Tokenizer vocabulary and mappings
├── stats.json       # Dataset statistics
└── data_config.json # Preprocessing configuration

Data Splits

Split	Percentage	Approximate Sequences
Train	90%	~90,000-900,000
Validation	5%	~5,000-50,000
Test	5%	~5,000-50,000

Exact numbers depend on source corpus size and configuration.

Dataset Creation

Curation Rationale

This dataset was created to enable training of efficient Czech language models that can:

Run on consumer-grade hardware (unlike large multilingual models)
Generate coherent Czech text with proper morphology and syntax
Serve as a foundation for domain-specific fine-tuning
Support Czech NLP research with accessible model sizes
Provide educational resources for learning about language models

Source Data

Data Collection and Processing

The dataset was created using a standardized pipeline (see prepare_data.py):

File Collection
- Recursive scanning of text files (.txt, .md, .rst, .py, .js, .html, .css, .json, .xml, .csv, .log, .c, .cpp, .java)
- Collection from multiple Czech text sources
Text Cleaning
- URL removal using regex patterns
- Email address removal
- Whitespace normalization (multiple spaces → single space)
- Short line filtering (minimum 10 characters)
- Deduplication of repeated content
Tokenization
- Character-level tokenization (configurable)
- Special tokens: <pad>, <unk>, <bos>, <eos>
- Vocabulary construction with minimum frequency threshold
- Default vocab size: 10,000 tokens
Sequence Creation
- Overlapping sequences with configurable stride
- Default: seq_len=512, stride=256
- Each sequence is seq_len + 1 tokens (513 for default)
- Ensures context preservation across sequences
Dataset Splitting
- Stratified split: 90% train, 5% validation, 5% test
- Random shuffling with fixed seed (42) for reproducibility

Who are the source data producers?

The source data comes from publicly available Czech text sources:

Czech Wikipedia articles (licensed under CC BY-SA)
Public domain Czech literature (classical authors)
Czech news websites (where redistribution is permitted)
Czech technical documentation (open-source projects)
Czech blogs and forums (publicly accessible)

All sources respect copyright laws and licensing requirements. No personal or private communications are included.

Annotations

This dataset contains no additional annotations beyond tokenization. It is designed for unsupervised language modeling.

Personal and Sensitive Information

Efforts have been made to remove personal information:

Email addresses: Automatically removed during preprocessing
URLs: Automatically removed during preprocessing
PII screening: Basic filtering applied

However, as with any web-scraped corpus, complete removal of personal information cannot be guaranteed. Users should be aware that residual personal information may exist and should implement additional safeguards for sensitive applications.

Bias, Risks, and Limitations

Known Limitations

Technical Limitations:

Character-level tokenization: Suboptimal for Czech morphology (consider BPE/WordPiece for production)
Fixed sequence length: Truncates long documents
Limited vocabulary coverage: 10,000 tokens may miss rare words
Limited coverage: Of Czech dialects and regional variations
Static dataset: Does not include recent events or information

Quality Limitations:

Variable text quality depending on source
Potential for noise from web-scraped content
Incomplete representation of all Czech language domains
May not capture spoken Czech or informal language adequately

Biases

The dataset may contain various biases:

Source Bias:

Overrepresentation of formal/written Czech vs. informal/spoken Czech
Skewed toward certain topics (e.g., technical, encyclopedic content)
Temporal bias reflecting when texts were written

Demographic Bias:

May reflect perspectives of source text authors
Potential underrepresentation of minority viewpoints
Geographic bias toward standard Czech vs. regional variants

Content Bias:

May perpetuate stereotypes present in source data
Potential political or ideological biases from source selection
Unequal representation across different subject domains

Citation

If you use this dataset, please cite:

BibTeX:

@misc{slim_cz_v1_dataset,
  title={SLiM-CZ-V1 Czech Text Corpus},
  author={Filip Sedivy},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/filipsedivy/SLiM-CZ-V1}}
}

APA:

Filip Sedivy. (2025). SLiM-CZ-V1 Czech Text Corpus. Hugging Face. https://huggingface.co/datasets/filipsedivy/SLiM-CZ-V1

Glossary

SLiM-CZ-V1: Slavic Linguistic integrated Micro-model for Czechia
Autoregressive: Model predicts next token based on previous tokens
Sequence Length (seq_len): Number of input tokens in each training sequence
Stride: Overlap between consecutive sequences (prevents context loss)
Token: Basic unit of text (character in this implementation)
Vocabulary Size: Number of unique tokens the model can represent
Character-level tokenization: Each character is a separate token (simpler but less efficient than BPE)

More Information

Dataset Statistics

Sequence Format: List of lists (no keys, just token IDs)
Sequence Length: seq_len + 1 tokens (default: 513)
Vocabulary Size: Configurable (default: 10,000)
Tokenization: Character-level (each character = 1 token)
Total Tokens: ~100M-1B (depending on source corpus)
Languages: Czech only
File Format: JSON (plain lists)

Quality Assurance

The dataset undergoes several quality checks:

Duplicate detection and removal
Minimum line length filtering (10 characters)
Character encoding validation (UTF-8)
Token frequency analysis
Sequence length verification (all sequences are seq_len + 1)
Split integrity checking

Dataset Card Contact

For questions, issues, or contributions:

GitHub Issues: https://github.com/filipsedivy/SLiM-CZ-V1/issues
Hugging Face Discussions: https://huggingface.co/datasets/filipsedivy/SLiM-CZ-V1/discussions

Downloads last month: 31