heymeowcat/sinhala-commons

Sinhala Commons

The flagship open corpus for Sinhala language technology: 42.35 million sentences gathered exhaustively from every publicly licensed Sinhala source we could find, cleaned with a documented pipeline, and released with a full datasheet so every byte is auditable.

Sinhala (සිංහල, sin_Latn/si) is spoken by ~17 million people but has a severe public-data deficit: the largest "open" datasets are unlicensed scrapes, and the licensed ones are news wire or translationese. This project exists to close that gap legally and cleanly — conversation register included, not just encyclopedic prose.

Configs

config	rows	schema	register	what it is
web	41,627,384	text, source, register	mixed: forum/blog/crawl/letters	the everyday-language bulk: cleaned crawl corpora + domain-tagged sentences + anonymized official letters
wiki	353,411	text, source, register	encyclopedic	full Sinhala Wikipedia (namespace 0, templates/references stripped)
conversations	279,582	turns (list of {speaker, text}), source, register	spoken/conversational	subtitle-dialogue threads, instruction/QA turns
parallel	74,602	si, pair_lang, pair_text	formal + colloquial	Sinhala↔English and Sinhala↔Tamil aligned pairs
romanized	13,686	text, source, register	Singlish (Latin-script)	romanized Sinhala (Singlish) chat sentences — kept out of the Sinhala-script configs by design; useful for transliteration work. Unfiltered by LID, so some pure-English rows remain (bonus config, v1)

News text is deliberately excluded by default (opt-in; see News, why excluded below). A social config is planned once a licensed social source exists — today none does, and we do not launder unlicensed scrapes into "open" datasets.

Example rows

web — everyday Sinhala · 41,627,384 rows

register	text
Conversational & Informal	ඒ මදිවට දැන් වංක භාවයත් පෙන්නන්නවා
Conversational & Informal	ඒ එන්නෙ හැමදාමත් ටීචර්ව මගෙ ළඟ තියාගන්න කියලා
letters	[යවන්නාගේ නම] ආණ්ඩුකාර ලේකම් [ලිපිනය] පිටපත : සභාපති, ප්‍රාදේශීය සභාව, [ලිපිනය]
formal	ඔයාගෙන් අහන්නේ නැතුවම ඔයාට ගිය අවුරුද්දේ බියර් එකක් අරන් දුන්නා!

wiki — encyclopedic · 353,411 rows

register	text
encyclopedic	එය ඕනෑම නිල පනතක් මත මූලික අයිතිවාසිකම් නඩු පැවරීමකට යටත් කිරීම මගින් ඕනෑම ජනාධිපතිවරයෙකුගේ මුක්තිය කප්පාදු කරන ලදී.
encyclopedic	පිරිවර සමග රන් ලියවැල ලෙසට බිසව්

conversations — dialogue threads · 279,582 rows

නියම උපදෙසක්, පින්කි. || සුභ උදෑසනක්. || නීතිඥ, පින්කස් සයන.
අපේ වයසෙ විතර ළමයි කණ්ඩායමක්, එයාල එක්ක ගෑණු ළමෙක් ඉන්නව. || කළු කොන්ඩේ. || ඔයා දන්නවද දෙයක්?

Each row's turns is a list of {speaker, text} objects.

parallel — si↔en and si↔ta pairs · 74,602 rows

si	pair_lang	pair_text
බස් එක මෙතන විනාඩි 20ක් නවත්තනවා	en	the bus stops here for 20 minutes
අන්තිමට මැරුණු කෙනා තාම ඇතුලේ	en	the dead last is still inside
එතුමා දන්නෙහිද?	ta	அவர் அறிவாரா?

romanized — Singlish (Latin-script) · 13,686 rows

register	text
conversational	Respect Gerhold oya la lankawe hodama sim eka wenawa mai super
conversational	Oka deactivated කරන්නෙ කෝමද
conversational	මම සිග්නල් කණුව ළඟටම ගිහින් speed එක බැලුවා..ඒත් ඒ ගානමයි.ඕයි..

Quickstart

from datasets import load_dataset

ds = load_dataset("heymeowcat/sinhala-commons", "web")
for row in ds["train"].shuffle(seed=0).select(range(5)):
    print(row["register"], "|", row["text"][:120])

conv = load_dataset("heymeowcat/sinhala-commons", "conversations")
print(conv["train"][0]["turns"])

Processing (applied identically to every source)

1. Unicode: NFC-normalized; ZWJ (U+200D) and ZWNJ (U+200C) are preserved — they carry orthographic meaning in Sinhala script.
2. PII masking: emails → <email>, URLs → <url>, phone numbers → <phone>; no raw contact data ships.
3. Quality filter: length bounds, degenerate-repetition guards (character 3-grams, repeated-token ratio).
4. Language ID: fastText lid.176 (Sinhala top-1 ≥ 0.30) with a Sinhala-script-ratio fallback.
5. Global exact dedup across all configs (first occurrence wins, processing order: web-bulk → conversations → wiki → remaining sources), with a small repeat-cap for legitimately repetitive short lines.
6. Splits: deterministic 99.5/0.5 hash split on content — stable across rebuilds.

Per-source kept/dropped counts ship in stats.json.

Datasheet

Source & license table

source	config	license	notes
Minuri/diverse_sinhala_dataset	web	CC-BY-SA-4.0	7.7M domain-tagged high-confidence sentences
9wimu9/sinhala_dataset_sanitized	web (+subtitle slice → conversations)	crawl-derived¹	mC4 slice ODC-BY, OSCAR slice CC0; news/FLORES files excluded
OPUS OpenSubtitles v2018 (si)	conversations	OPUS research terms	subtitle lines grouped into dialogue threads
OPUS QED v2.0a (si)	web	MIT	small educational sentences
Sinhala Wikipedia dump	wiki	CC-BY-SA-4.0	namespace 0 only
NLPC-UOM code-mixed/code-switched	romanized	MIT	10K+ si-en annotated Singlish (romanized) sentences
NLPC-UOM anonymized letters	web	CC-BY-SA-4.0	<1K official letters, anonymized at source
NLPC-UOM Sinhala-Tamil aligned	parallel	MIT	parliament-register pairs
zaanind si-en parallel	parallel	GPL	81K colloquial pairs; this slice remains GPL inside the aggregate
Suchinthana QA-Translate	parallel	MIT	1K QA pairs
Suchinthana Dolly-15k-Sinhala	conversations	CC-BY-SA-3.0	instruction turns; machine-translated (flagged)

¹ The 9wimu9/sinhala_dataset_sanitized repo carries no license tag. Its components are mC4 (ODC-BY 1.0), OSCAR (CC0-1.0), and blog/local crawl slices of the same provenance class that mC4/OSCAR themselves were redistributed under. We anchor each slice to that lineage and state it here; if you object to a slice, the per-file structure lets you drop it.

What was searched and rejected (summary of the full audit)

The project searched all of HuggingFace (language-tag si, org listings, 13+ query families), OPUS (file-level HEAD probes), Wikimedia dumps, archive.org (including a CC-license filter), DOAJ, GitHub, and Common Voice. Rejections fell into five buckets:

• No license — Navanjana/Sinhala-Chat (1.2M rows!), SOLD/SemiSOLD (Twitter), most "Sinhala conversations" datasets. We will not redistribute unlicensed text, however large the temptation.
• Non-commercial — FLORES, Khan Academy Sinhala, XLSum: excluded to keep the corpus commercially usable.
• Synthetic/MT contamination — Google-translated customer-support sets, LLM-generated ad/chat data.
• Licensed but not worth shipping — Sachin-Hansaka/SQAD (CC-BY-SA-4.0, login-gated): only 73K of 110K rows are usable QA pairs, answers are span-short (median 13 chars), and its context passages duplicate the wiki register. Revisit if users ask.
• Verified absence — OPUS Tatoeba/WikiMatrix/GlobalVoices/Ubuntu/ GNOME have no Sinhala (probed); Common Voice Sinhala does not exist.

Known limitations & biases

• Register: true social-media Sinhala (Facebook comments, YouTube, TikTok) is absent by design — every such platform is all-rights-reserved/ToS-protected. The closest proxies we ship are subtitle dialogue, romanized code-mixed chat sentences, and colloquial parallel pairs.
• Dialect/variety: formal written Sinhala dominates; colloquial (spoken-style) text is minority-weight. Regional dialect coverage is unknown.
• Machine translation: two small sources are MT-derived (Dolly-si, noted above); they are a drop in the 42.35M-row bucket.
• Names: person names (politicians, historical figures) appear in wiki/news-register text as encyclopedic fact. Downstream keyboard/LM users who want name suppression should handle it at the model level (see CleanBoard's name-filtering approach) rather than expecting it from raw text.
• Crawl-era drift: crawl slices skew toward text that existed online 2015–2024; very recent coinages are underrepresented.

Suggested uses

• Language-model pretraining/finetuning for Sinhala
• Keyboard/IME language models (n-gram, neural autocomplete) — this corpus is the training data behind CleanBoard's next models
• Machine translation (parallel config)
• Register-conditional generation (filter by register)

Not intended for

• Benchmarking against test sets that may leak: crawl-derived corpora can overlap public eval sets (FLORES, SinQuAD); decontaminate before reporting numbers.

Building from source

python3 scripts/fetch.py        # ~8 GB into raw/ (resumable)
python3 scripts/process.py      # raw/ -> dist/dataset/*.parquet
python3 tests/test_process.py   # synthetic end-to-end regression test

Every decision — including the ~40 rejected sources and why — lives in scripts/sources.py. Issues and additions welcome via PR.

Citation

@misc{sinhala-commons,
  title = {Sinhala Commons: an audited open corpus for Sinhala},
  author = {Vidura Fernando},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/heymeowcat/sinhala-commons}},
  note = {See LICENSES.md for per-source licenses}
}