Qor Af-Soomaali
Qor Af-Soomaali — write Somali — is the data program behind our evaluation research: a community platform where Somali speakers write, translate, transcribe, and validate the language, and a versioned open dataset built from what they make. It exists because safety evaluation has to be built out of language, and for Somali that language had to be created, not scraped.
The corpus today
Counts update live from the platform. A sentence is counted once two independent community validators accept it; releases additionally require linguist verification.
How a sentence earns its place
- A contributor writes.
Before their first sentence, every contributor explicitly consents to the open license and chooses how to be credited: by name, by pseudonym, or anonymously. The item records its mode, register, domain, and the contributor’s dialect from the moment it is written.
- Two peers validate.
Two independent community members judge each submission. Two approvals accept it; two rejections reject it. Nobody validates their own work, and the platform enforces one vote per person per item.
- Disagreements escalate.
A split vote escalates automatically, and a trusted reviewer’s vote settles it. The full validation record stays attached to the item.
- A linguist verifies.
Community-accepted items pass to linguist reviewers for final sign-off. Only verified items enter official releases.
- A versioned release ships.
Verified items are exported with their provenance into a versioned dataset on Hugging Face, with an auto-generated dataset card and a credits file honoring each contributor’s choice.
Four ways to contribute
- Qor · WriteContributors respond to prompts about everyday life — stories, instructions, dialogues, opinions — or write on a topic of their own. These are the registers of Somali that never made it onto the web.
- Turjun · TranslateShort sentences move from English into Somali, producing the parallel data that makes like-for-like evaluation possible.
- Guuri · TranscribeOpenly licensed and public-domain printed Somali is typed up, turning paper heritage into digital text.
- Hubi · ValidateContributors review each other’s work, asking one question: is this correct, natural Somali? It takes seconds and can be done on a phone.
What a released record looks like
Field names below are the dataset’s actual schema. The values shown are illustrative; every real record carries a known author, a consent record, and its validation history.
{
"id": "…",
"text_so": "…", // the Somali text
"text_en": null, // source sentence, translation mode only
"meaning_en": null, // meaning gloss, proverb mode only
"mode": "write", // write | translate | transcribe | proverb
"register": "everyday",
"sector": "health", // one of nine domains
"topic": null, // contributor-stated topic, free-writing only
"dialect": "maxaa_tiri", // the CONTRIBUTOR’s spoken variety, not the text’s orthography
"verified": true, // linguist-verified tier
"license": "CC BY-SA 4.0",
"created_at": "2026-08-09"
}Consent and governance
Nobody’s words enter the corpus without their knowledge. Consent is collected before contribution, not after; credit is the contributor’s choice; and account deletion never deletes the consent record attached to already-released work. All data is released under CC BY-SA 4.0 in versioned datasets, so researchers, developers, and the Somali community itself can build on it without restriction. The full collection policy and dataset documentation live on the dataset card and in the platform source, which is open so that any language community can deploy it.
Release history
| Version | Date | Verified records | Notes |
|---|---|---|---|
| v0.2.1 | 2026-08-09 | 268 | Current release · 2,282 verified sentences |
| v0.2.0 | 2026-08-04 | 235 | |
| v0.1.2 | 2026-07-30 | 124 | |
| v0.1.1 | 2026-07-30 | 124 | |
| v0.1.0 | 2026-07-30 | 124 | First public release |
Working with the program
Universities, schools, and organizations that want to bring their community onto Qor — or take the platform to another language — should get in touch.