superAVTR/wikipedia_en_512_for_pretraining

Cleaned Wikipedia 512 Pretraining Dataset

Dataset Description

This dataset is a cleaned version of the Hugging Face dataset lucadiliello/wikipedia_512_pretraining.

The original dataset contains English Wikipedia text prepared for language-model pretraining. This derivative version applies additional filtering intended to remove malformed, duplicated, or incomplete training samples while otherwise preserving the original text.

Source

Original dataset:

• lucadiliello/wikipedia_512_pretraining
• Source language: English
• Source format: Parquet
• Original content: Wikipedia text prepared for pretraining

Additional Cleaning Done

The cleaning pipeline performs the following operations:

• normalizes whitespace and removes invalid control characters;
• removes repeated text at the beginning of a document when an immediately duplicated prefix is detected;
• removes repeated text at the end of a document when an immediately duplicated suffix is detected;
• removes duplicated final sentences;
• removes unmatched or clearly extraneous trailing characters;
• removes incomplete text fragments at the beginning of a document when a valid sentence boundary can be identified;
• removes incomplete final sentences;
• discards documents that do not contain a valid completed sentence after cleaning;
• removes exact duplicate documents using a SHA-1 hash of the final cleaned text.

The cleaner is intentionally conservative and does not attempt semantic deduplication or rewriting of the document body.

Dataset Structure

Each JSONL record contains:

{
  "text": "Cleaned document text.",
  "num_words": 3,
  "num_chars": 22,
  "document_hash": "SHA-1 hash of the cleaned document"
}

Fields:

• text — final cleaned document text;
• num_words — number of whitespace-separated words in the cleaned document;
• num_chars — number of characters in the cleaned document;
• document_hash — SHA-1 hash calculated from the final cleaned text.

Intended Use

The dataset is intended primarily for language-model pretraining and experiments involving small and medium-sized language models. The filtering process improves structural cleanliness but does not verify the factual accuracy of individual Wikipedia passages and does not perform semantic quality filtering.

Limitations

This dataset inherits the content, biases, factual errors, and other limitations of its original source. The filtering rules are heuristic. Some malformed documents may remain, and some valid documents may be removed by the cleaning process.