PH Pretrain 03 — Filipino/Tagalog Web Corpus (ph-pretrain-03)

The recommended dataset for Filipino / Tagalog language-model pretraining and continued pretraining (CPT).

A ~1-billion-token, Tagalog-forward web corpus, cleaned, deduplicated, and quality-scored by the jpaulpoliquit/pretraining refinery.

• 1,933,988 documents · ≈ 1.0 billion tokens (Sailor2-1B, calibrated)
• ~97.7% Tagalog web (FineWeb2 + CC100), with regional + news + Wikipedia in the long tail
• ~3.41 GB to download (Parquet, zstd) · ~15.1 GB uncompressed
• Higher quality share than Wikipedia-dominated corpora: 59% bucket A

Which PH dataset should I use?

• Use this one (ph-pretrain-03) if your goal is a model that is good at Filipino/Tagalog — it is web-text, Tagalog-forward, and quality-filtered.
• Use jpaulpoliquit/ph-pretrain only if you specifically want mass Cebuano/Waray Wikipedia (regional coverage). That set is ~95% bot-templated ceb/war Wikipedia and is not the recommended default for Filipino modeling.

⚠️ Licensing: ~49% of rows are train-only / Common-Crawl-derived and are not redistribution-ready. The dataset is fine to train on; if you want to re-publish it, first filter to redistributable rows (source_type == "wikimedia_dump" / license == "odc-by-1.0"). See Licensing.

Release manifest (source of truth): metadata/release_ph-pretrain-03.json.

Composition

By source (documents)

Source	Documents	Share	License tier
fineweb2_tl_latn (Tagalog web, FineWeb2)	946,672	48.95%	ODC-BY (redistributable)
cc100_tl_documents (Tagalog, CC100)	943,593	48.79%	Common Crawl terms (train-only)
halohalo_combined (Filipino curated)	30,946	1.60%	ODC-BY (redistributable)
wikimedia_ceb (Cebuano Wikipedia)	12,112	0.63%	CC-BY-SA-4.0 (redistributable)
rappler_news	414	0.02%	Fair-use research (train-only)
gma_news	184	0.01%	Fair-use research (train-only)
philstar_psn	67	<0.01%	Fair-use research (train-only)

Tagalog web (FineWeb2 + CC100) = 97.74% of documents.

By language bucket

Bucket	Documents
filipino_casual	963,676
filipino_formal	910,395
taglish (Tagalog–English code-mix)	47,241
cebuano	12,187
hiligaynon / ilocano / bicolano / other	<500 each

The dominant detected language is Tagalog (tl); register/language_bucket are heuristic refinery labels.

Quality bands

Heuristic quality_score (0–1) and A/B/C/D quality_bucket:

Bucket	Documents	Share
A	1,146,176	59.27%
B	784,809	40.58%
C	2,987	0.15%
D	16	<0.01%

Splits

Split	Documents
train	1,895,405
validation	19,462
test	19,121

Per-document holdouts assigned at mix time (~1% each). Convenience holdouts, not a contamination-controlled benchmark.

Size & tokens

Metric	Value
Documents	1,933,988
Whitespace words	~772 M
Tokens (Sailor2-1B tokenizer)	1,003,029,302
Download size (Parquet)	~3.41 GB
Uncompressed	~15.1 GB

The token count is calibrated: the refinery tokenized a 4,000-document sample with sail/Sailor2-1B (calibrated 1,003,029,302 vs. heuristic 1,002,300,602). This release meets the web-1b scale gate (≥ 1B tokens).

Usage

from datasets import load_dataset

# Full dataset (~3.41 GB download)
ds = load_dataset("jpaulpoliquit/ph-pretrain-03")
row = ds["train"][0]
print(row["language"], row["register"], row["text"][:300])

# Stream without downloading everything
stream = load_dataset("jpaulpoliquit/ph-pretrain-03", split="train", streaming=True)
for row in stream:
    text = row["text"]
    break

Carve a redistributable-only subset, or keep only high-quality Tagalog:

redistributable = ds["train"].filter(lambda r: r["license"] in ("odc-by-1.0", "cc-by-sa-4.0"))
hq_tagalog = ds["train"].filter(lambda r: r["language"] == "tl" and r["quality_bucket"] == "A")

For pretraining you usually only need text (plus language / quality_bucket / license for sampling and compliance).

Schema

Flat Parquet at data/{train,validation,test}/part-*.parquet.

Column	Type	Description
id	string	Stable document id
text	string	Full document body (UTF-8)
text_truncated	string	Preview (≤ 2048 chars)
url	string	Source URL when available
date	string	created_date if set, else crawl_date
dump	string	Upstream dataset id (e.g. HuggingFaceFW/fineweb-2)
file_path	string	Provenance pointer into the refinery release
detected_lang / language	string	Language-ID code
word_count	int64	Document word count
title	string	Title, slug, or first line
source	string	Refinery source_name
token_count	int64	Token count (tokenizer when present, else ~1.3× words)
content_hash	string	text_sha256 (dedup key)
crawled_at	string	Refinery ingest timestamp
register	string	Editorial register label
language_bucket	string	Coarse language/register bucket
language_confidence	float64	Language-ID confidence (may be sparse)
quality_score	float64	0–1 heuristic quality
quality_bucket	string	A / B / C / D
license	string	Authoritative per-row license
source_type	string	hf_curated, wikimedia_dump, or news_formal
release_version	string	ph-pretrain-03
split	string	train / validation / test
metadata_json	string	Remaining refinery fields as JSON

How it was built

Pipeline (full source: jpaulpoliquit/pretraining):

1. Ingest FineWeb2 (tl), CC100 (tl), curated Filipino, regional Wikipedia, and a little PH news.
2. Language ID + register/bucket assignment (Tagalog, Taglish, formal/casual, regional).
3. Quality filtering & scoring, with low-quality (C) capping.
4. Exact dedup by content hash — 632,106 duplicate documents dropped during this build (shard-local + global scope).
5. Mix & split to the web-1b scale contract (≥ 1B tokens) with per-source token budgets.

Licensing

This corpus mixes licenses; the per-row license column is authoritative.

License	Documents	Share	Redistributable?
odc-by-1.0 (FineWeb2, halohalo)	977,618	50.55%	✅ Yes (attribution)
common_crawl_terms_research (CC100)	943,593	48.79%	❌ Train-only
cc-by-sa-4.0 (Wikimedia)	12,112	0.63%	✅ Yes (attribution + share-alike)
fair_use_research_only (news)	665	0.03%	❌ Train-only

Legal tiers: open_redistributable = 989,730 docs (51.2%) · train_only_no_redistribution = 944,258 docs (48.8%).

The dataset is tagged license: other because ~49% of it is not redistribution-cleared. You may train on the full set for research; to re-publish, filter to the redistributable tiers (license in {odc-by-1.0, cc-by-sa-4.0}). Underlying web text derives from Common Crawl and remains subject to the original publishers' rights. You are responsible for your own license compliance.

Intended use

• Recommended default for continued pretraining / domain adaptation of LMs targeting Filipino / Tagalog (incl. Taglish).
• Tokenizer training and Filipino web-text analysis.
• Pair with jpaulpoliquit/ph-pretrain if you also need heavy regional (Cebuano/Waray) coverage.

Limitations & biases

• Web-crawl noise. Despite filtering, expect boilerplate, SEO text, and occasional adult/spam residue typical of Common Crawl derivatives.
• Train-only licensing on ~49% (see Licensing) — not redistribution-ready as-is.
• Heuristic labels. register, language_bucket, and quality_score are automatic, not gold; Taglish and closely related languages are sometimes mislabeled.
• Minor non-Tagalog leakage (en, ceb, etc.) remains in the tl buckets.
• Train/val/test are random holdouts, not a leakage-controlled benchmark.

Provenance & reproducibility

• Manifest (source of truth): metadata/release_ph-pretrain-03.json
• Pipeline & build scripts: github.com/jpaulpoliquit/pretraining
• Release version: ph-pretrain-03 · scale gate: web-1b · mix recipe SHA256: c09f706af792feba…

Citation

@misc{ph_pretrain_03,
  title        = {PH Pretrain 03: A Filipino/Tagalog Web Pretraining Corpus},
  author       = {Poliquit, John Paul},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/jpaulpoliquit/ph-pretrain-03}
}

Please also cite the upstream sources: FineWeb2 (ODC-BY-1.0), CC100, and Wikimedia (CC-BY-SA-4.0).