Yxanul/Romanian-Wiki

Romanian Wikipedia - Processed Dataset

This is a processed version of the Romanian Wikipedia subset from the FineWiki dataset, optimized for language model training and analysis. The dataset has been filtered to essential fields and enriched with token counts using tiktoken's cl100k_base encoding.

Dataset Overview

• Total Pages: 493,462
• Total Size: ~669 MB (compressed parquet)
• Language: Romanian (ro)
• Total Tokens: ~395 million tokens (cl100k_base encoding)
• Source: FineWiki - An updated and better extracted version of the wikimedia/Wikipedia dataset originally released in 2023

Dataset Structure

Data Instances

Example from the Romanian subset (values truncated for readability):

{
  "id": "rowiki/3217840",
  "title": "Melba (film)",
  "url": "https://ro.wikipedia.org/wiki/Melba_(film)",
  "date_modified": "2023-08-15T10:22:31Z",
  "text": "# Melba (film)\nMelba este un film biografic muzical britanic din 1953 regizat de Lewis Milestone...",
  "token_count": 2145
}

Data Fields

• id (string): dataset-unique identifier; format: rowiki/<page_id>
• title (string): article title
• url (string): canonical article URL
• date_modified (string): ISO-8601 timestamp of the last page revision
• text (string): cleaned, structured article text preserving headings, lists, code/pre blocks, tables and math. Has markdown formatting (headings, tables, lists)
• token_count (int64): number of tokens in the text field, calculated using tiktoken's cl100k_base encoding (used by GPT-4 and similar models)

Tokenization

Token counts are computed using tiktoken with the cl100k_base encoding, which is the same tokenizer used by:

• GPT-4
• GPT-3.5-turbo
• text-embedding-ada-002

This makes the dataset particularly useful for training or fine-tuning models compatible with OpenAI's tokenization scheme.

Processing Details

This dataset was created from the original FineWiki Romanian subset by:

1. Filtering from 14 columns down to 6 essential fields
2. Computing token counts for each article's text using tiktoken (cl100k_base)
3. Processing in batches of 10,000 rows for efficient computation
4. Saving as compressed parquet files with snappy compression

Files

• 000_00000_processed.parquet: 249,533 articles (~327 MB)
• 000_00001_processed.parquet: 243,929 articles (~342 MB)

Use Cases

• Training or fine-tuning Romanian language models
• Token budget analysis and dataset planning
• Information retrieval and semantic search
• Question answering systems
• Text summarization and generation tasks

Citation Information

@dataset{penedo2025finewiki,
  author    = {Guilherme Penedo},
  title     = {FineWiki},
  year      = {2025},
  publisher = {Hugging Face Datasets},
  url       = {https://huggingface.co/datasets/HuggingFaceFW/finewiki},
  urldate   = {2025-10-20},
  note      = {Source: Wikimedia Enterprise Snapshot API (https://api.enterprise.wikimedia.com/v2/snapshots). Text licensed under CC BY-SA 4.0 with attribution to Wikipedia contributors.}
}

License

The text content is licensed under CC BY-SA 4.0 with attribution to Wikipedia contributors, as per the original Wikipedia content license.

Dataset Creator

Processed and uploaded by Yxanul