metadata
license: mit
Dataset Card for Cleaned & Chunked WikiText Corpus
📌 Dataset Description
This dataset is a cleaned and chunked version of WikiText-style corpus, designed for language model pretraining and evaluation.
The preprocessing pipeline follows a lightweight data curation paradigm:
cleaning → deduplication → normalization → token-aware chunking
Specifically, the dataset includes:
- Removal of special and noisy characters
- Cleaning of formatting artifacts (e.g., HTML tags, irregular symbols)
- Deduplication to reduce redundant or highly similar text samples
- Text normalization (whitespace, encoding, etc.)
- Chunking of long documents into smaller segments suitable for model training
Each sample corresponds to a cleaned text chunk, rather than a full original document.
📊 Data Structure
Each sample in the dataset follows this structure:
{
"uid": "string",
"content": "string",
"meta_data": {
"index": "int",
"total": "int",
"length": "int"
}
}