rsateam/sr-bs-hr-clean-text

SR/BS/HR Clean Text Corpus

High-quality, deduplicated text corpus for Serbian, Bosnian, and Croatian

RSA Team • GitHub • Website

Overview

This dataset provides a carefully curated and cleaned text corpus for South Slavic languages, specifically designed to address quality issues found in existing corpora like OSCAR and CC100. It serves as a foundation for training language models, tokenizers, and conducting linguistic research on Balkan languages.

Why This Dataset?

Existing sr/bs/hr corpora often suffer from:

Problem	Our Solution
HTML fragments & noise	Aggressive cleaning pipeline
Poor deduplication	SHA256 + MinHash (>95% dedup rate)
Mixed languages	Source-based labeling + FastText validation
Unclear sources	Full provenance tracking

Dataset Statistics

Metric	Value
Total examples	641,186
Dataset size	2.67 GB
Download size	1.37 GB
Source	Wikipedia

Splits

Split	Examples	Size
train	512,948	2.14 GB
validation	64,116	266 MB
test	64,122	269 MB

Languages

Language	ISO Code	Source
Serbian	sr	sr.wikipedia.org
Bosnian	bs	bs.wikipedia.org
Croatian	hr	hr.wikipedia.org

Dataset Structure

Data Format

{
  "id": "550e8400-e29b-41d4-a716-446655440000",
  "title": "Article Title",
  "text": "Full cleaned article text...",
  "language": "sr",
  "source": "sr.wikipedia.org",
  "domain": "wiki",
  "date": null,
  "url": "https://sr.wikipedia.org/wiki/..."
}

Fields

Field	Type	Description
id	string	Unique identifier
title	string	Article title
text	string	Cleaned textual content
language	string	Language code (sr/bs/hr)
source	string	Source domain
domain	string	Content type (wiki)
date	null	Publication date (not available for wiki)
url	string	Original Wikipedia URL

Data Processing Pipeline

┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│  Collection │ -> │  Cleaning   │ -> │   Dedup     │ -> │  Lang ID    │ -> │  Filtering  │
│  Wikipedia  │    │  normalize  │    │  MinHash    │    │  FastText   │    │  quality    │
└─────────────┘    └─────────────┘    └─────────────┘    └─────────────┘    └─────────────┘

Processing Steps

1. Collection: Wikipedia dumps from bs/hr/sr Wikipedia
2. Cleaning: Markup removal, Unicode normalization (NFC), whitespace normalization
3. Deduplication: SHA256 exact matching + MinHash near-duplicate detection (90% threshold)
4. Language ID: Source-based labeling with FastText validation
5. Quality Filtering: Length constraints, language confidence >0.90

Usage

Loading the Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("rsateam/sr-bs-hr-clean-text")

# Load specific split
train = load_dataset("rsateam/sr-bs-hr-clean-text", split="train")

# Filter by language
serbian = dataset["train"].filter(lambda x: x["language"] == "sr")

Streaming

from datasets import load_dataset

dataset = load_dataset(
    "rsateam/sr-bs-hr-clean-text",
    split="train",
    streaming=True
)

for example in dataset:
    print(example["title"], "-", example["text"][:100])

Training a Tokenizer

from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from datasets import load_dataset

dataset = load_dataset("rsateam/sr-bs-hr-clean-text", split="train")

def batch_iterator(batch_size=1000):
    for i in range(0, len(dataset), batch_size):
        yield dataset[i:i+batch_size]["text"]

tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
trainer = BpeTrainer(
    vocab_size=32000,
    special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
)
tokenizer.train_from_iterator(batch_iterator(), trainer=trainer)

Supported Tasks

• Language Model Pretraining: Foundation for training or continued pretraining of LLMs
• Tokenizer Training: Clean text for BPE/WordPiece/Unigram tokenizer training
• Word Embeddings: Training Word2Vec, FastText, or similar embeddings
• Linguistic Research: Analysis of Serbian, Bosnian, and Croatian texts

Considerations

Ethical Considerations

• Data sourced from Wikipedia under CC-BY-SA license
• No personally identifiable information (PII)
• Encyclopedic content with neutral point of view

Limitations

• Single source (Wikipedia) — encyclopedic style only
• Some topics may be underrepresented
• Article length varies significantly

License

This dataset is released under CC-BY-SA-4.0, consistent with Wikipedia's licensing.

Citation

@dataset{rsateam_clean_text_2026,
  title={SR/BS/HR Clean Text Corpus},
  author={RSA Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/rsateam/sr-bs-hr-clean-text},
  note={High-quality deduplicated corpus for Serbian, Bosnian, and Croatian}
}

Future Plans

We plan to expand this dataset with additional sources:

• News portals (klix.ba, index.hr, blic.rs, etc.)
• Government and public institution documents
• Other curated text sources

Contributing

We welcome contributions! For suggestions, bug reports, or improvements:

• Open an issue on GitHub
• Email: office@rsateam.com

---

RSA Team — Building bridges between languages and AI, one dataset at a time.