README.md

metadata

language:
  - id
license: cc0-1.0
task_categories:
  - text-generation
pretty_name: IndoCleanCorpus

IndoCleanCorpus

A large-scale cleaned Indonesian text corpus, built from multiple open sources and processed through a multi-stage quality filtering and deduplication pipeline.

Source Datasets

Dataset	Description
Lyon28/Corpus-Indonesia	General Indonesian text corpus
DamarJati/indocorpus-sastra	Indonesian literary corpus
indonesian-nlp/wikipedia-id	Indonesian Wikipedia dump

Processing Pipeline

The raw merged corpus of 19,814,163 documents was cleaned through the following sequential stages:

Stage	Remaining Docs	Removed
Raw input	19,814,163	—
Boilerplate filter	19,811,935	2,228
Length filter	14,575,984	5,235,951
Character ratio filter	14,377,694	198,290
Repetition filter	14,376,725	969
Language filter (Indonesian)	13,528,694	848,031
Exact deduplication	13,431,107	97,587
Near-deduplication	13,411,142	19,965
Final	13,411,132	6,403,031

Overall retention rate: 67.68% — roughly 1 in 3 documents were removed.
Pipeline runtime: ~6.7 hours (23,995 seconds).

Filter Descriptions

• Boilerplate filter — removes templated or auto-generated content (nav menus, footers, repeated headers)
• Length filter — removes documents that are too short or too long to be useful
• Character ratio filter — removes documents with abnormal ratios of punctuation, digits, or non-alphabetic characters
• Repetition filter — removes documents with excessive repeated n-grams or lines
• Language filter — retains only documents classified as Indonesian (id)
• Exact deduplication — removes byte-identical duplicate documents
• Near-deduplication — removes near-duplicate documents using MinHash LSH

Usage

from datasets import load_dataset

ds = load_dataset("AiRukua/IndoCleanCorpus")

Dataset Statistics

• Final document count: 13,411,132
• Format: Parquet (16 part files, ~2.2 GB total)
• Language: Indonesian (id)

License

Source datasets are openly licensed. This dataset is released under CC0 1.0.