README.md

metadata

language:
  - id
size_categories:
  - 1M<n<10M
task_categories:
  - text-generation
  - fill-mask
license: cc-by-sa-4.0
tags:
  - indonesian
  - pretrain
  - wikipedia
  - culturax
  - clean

AksaraLLM Pretrain Clean v1

Versi clean dari AksaraLLM/aksara-pretrain-id, ditambah Wikipedia-id (Nov 2023 fresh dump).

Changes dari aksara-pretrain-id v4

Fix	Before	After
Exact duplicate rows	48,009 (5.84%)	0
Train/val leakage	11.76%	0% (hash-based split)
NusaX [Bahasa X] prefix	5,388 rows contaminated	0 (stripped)
Malay rows labeled as Indonesian	~25%	≤ 5% (GlotLID P≥0.60 filter)
Gopher-style quality filter	No	Yes
URL blocklist (judi/spam)	No	Yes
Fresh Wikipedia-id 2023-11-01 dump	No	Yes (~665k articles)

Schema

{
  "text": string,       # document content
  "source": string      # one of: wikipedia-id, culturax-id, wikipedia-indonesia-topik,
                        # wikipedia-id-20231101, nusax-{ace,ban,bbc,bjn,bug,ind,jav,mad,min,nij,sun}
}

Methodology

1. Load train + validation dari aksara-pretrain-id.
2. Strip [Bahasa X] prefix dari rows NusaX.
3. Exact MD5 dedup.
4. Gopher-style filters:
   • Min 80 chars, avg words per line ≥ 3
   • Alphabetic char fraction ≥ 0.65
   • Fraction of lines ending with "…" ≤ 0.30
   • Nav boilerplate density ≤ 5%
   • URL blocklist (hand-curated: gambling, spam domains)
5. GlotLID language classifier (cis-lmu/glotlid):
   • For id-partition sources: keep only ind_Latn with P ≥ 0.60
   • For NusaX sources: keep all (FT may not support those labels)
6. MinHash near-dedup (Jaccard threshold 0.85, num_perm=128, 5-gram shingles).
7. Merge fresh Wikipedia-id (wikimedia/wikipedia 20231101.id), deduped vs existing.
8. Deterministic hash-based train/val split (last hex = "0" → val, no leakage possible).

Stats (v1.0 — April 2026)

Stage	Rows
Raw (aksara-pretrain-id train + val)	839,366
After exact MD5 dedup	789,368
After Gopher quality filter	770,310
After GlotLID language filter	758,642
+ Fresh Wikipedia-id (2023-11-01, deduped)	+104,473
Final total	863,115
→ train split	808,886
→ validation split	54,229

Estimated tokens (chars/3 heuristic):

• Train: ~500 M
• Validation: ~34 M

Language distribution (after filter):

• ind_Latn: 99.5% (754,973 / 758,642)
• jav_Latn, sun_Latn, min_Latn, ban_Latn, bjn_Latn, ace_Latn, bug_Latn, mad_Latn, bbc_Latn: ~350–400 each (= NusaX sentiment subsets, kept despite low confidence because source label is trusted)

Split

Split	Strategy
train	MD5(text)[0] != "0"
validation	MD5(text)[0] == "0" (~6% of data)

Karena deterministic by text hash, tidak mungkin ada leakage saat dataset di-update di masa depan: document yang sama akan selalu masuk ke split yang sama.

Licensing

IMPORTANT: Dataset ini adalah gabungan dari multiple sources dengan license yang berbeda:

Source	License
wikipedia-id + wikipedia-id-20231101 + wikipedia-indonesia-topik	CC-BY-SA 4.0
culturax-id	ODC-BY (dari OSCAR/mC4)
NusaX subsets	CC-BY-SA 4.0

Secara keseluruhan, license paling ketat adalah CC-BY-SA 4.0, dan dataset ini dirilis under that license. Pastikan kamu mematuhi terms untuk hilirnya.

Reproduce

Lihat scripts/clean_pretrain.py di repo AksaraLLM/aksara-data (cabang clean-v1).

pip install datasets fasttext-wheel datasketch "numpy<2.0"
wget https://huggingface.co/cis-lmu/glotlid/resolve/main/model.bin -O glotlid.bin
python scripts/clean_pretrain.py

Citation

Kalau pakai dataset ini, mohon cite:

@misc{aksarallm_pretrain_clean_v1,
  author       = {AksaraLLM Community},
  title        = {AksaraLLM Pretrain Clean v1},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/AksaraLLM/aksara-pretrain-clean-v1}}
}

Plus upstream sources:

• Nguyen et al. 2023 (CulturaX)
• Winata et al. 2023 (NusaX)
• Wikimedia Foundation (Wikipedia)