halohalo

Dataset Summary

halohalo is a Pretraining text corpus for Philippine languages, assembled from web-scraped data. It is compatible with Fineweb for LLM Pretraining.

Source Data

Derived from the following cleaned datasets:

Source	Documents
halo-hil	8,874
halo-tgl	6,589
halo-bcl	1,264

Each source dataset was cleaned using clean_halo.py to remove web boilerplate, navigation menus, markdown noise, HTML artifacts, and low-quality documents before being included here.

Processing

1. Cleaning (clean_halo.py) — strips boilerplate, HTML, markdown noise; filters documents with fewer than 30 words or less than 40% Latin characters
2. FineWeb formatting (prep_halohalo.py) — adds source, language, token_count, content_hash; deduplicates against existing documents using MD5 content hashing

Processing code is available at github.com/sapinsapin/halohalo.

Statistics

Metric	Value
Total documents	16,727
Total tokens	19,178,582
Avg tokens per document	1,146.6
Min tokens	30
Max tokens	10,552

Languages

Language	Documents	Word Count
hil	8,874	9,332,784
tgl	6,589	8,208,749
bcl	1,264	1,637,049
Total	16,727	19,178,582

Schema

Field	Type	Description
text	str	Cleaned document text
id	str	Unique document identifier
source	str	Source dataset name
language	str	ISO 639-3 language code
token_count	int	Whitespace-tokenized word count
content_hash	str	MD5 hash of text for deduplication
url	str	Source URL
date	str	Crawl date
dump	str	CommonCrawl dump identifier
title	str	Page title

Usage

from datasets import load_dataset

ds = load_dataset("sapinsapin/halohalo")
print(ds["train"][0])