updated link: discord.gg/kDNWDhHv

All are welcome. Please make this space your own, and add channels at will.

Our first task is to massively expand SMOL through community contribution. Anyone who contributes significant volunteer translations or post-edits will get on the Arxiv paper in the next refresh!

this is a space for grassroots collaboration. It doubles as a directory of speakers of such languages, so you can directly talk with and collaborate with community members.

Working on Low Resource Languages? Want to help with SMOL? join our new discord! discord.gg/YFTv7tkh

By the way, GATITOS has now officially moved to the SMOL Huggingface repo

Finally, if you are a speaker of any SMOL languages, please take a look at the data and tell me what you think. Despite the quality checks, I am sure that some of the deliveries have quality issues, and I would love to understand and/or fix any affected sources. We are in this together!

I would also like to thank FAIR for being an academic leader in open-sourcing work with low-resource languages, including NLLB and Flores. Thank you for helping make the academic community feel collaborative!

I would like to thank our native-language consultants and translators -- too numerous to name -- for their invaluable help along the way. Several entire languages in SMOL only exist because of volunteer contributions!

SMOL also provides factuality ratings for 671 documents, with well-researched justifications.

SMOL has two sub-sources: SMOL-Doc, a document-level set, and SMOL-Sent, a sentence-level source. They join the token-level GATITOS to hit at three levels of granularity!

And that’s just OOTB finetuning—we know that the community can think of more clever ways to train on SMOL. Multiway parallel data is tricky to deal with without overfitting.

Finetuning of Gemini 2.0 Flash on SMOL yields average improvements of about +4.0 ChrF, with some languages -- including Ewe, Kokborok, Manipuri, Ga, and Dombe -- seeing gains of over +20 ChrF.

SMOL comprises sentences and documents carefully selected for the biggest “Bang for Buck” ratio. It includes 6.1M translated tokens—and if you’ve been in this field a while you know that’s a lot!

😼SMOL DATA ALERT! 😼Anouncing SMOL, a professionally-translated dataset for 115 very low-resource languages! Paper: arxiv.org/pdf/2502.12301
Huggingface: huggingface.co/datasets/goo...

Is Dravidian negation illa (இல்லை etc.) cognate to Semitic *lā (لَا etc.?) there was a lot of trade in that region so it seems likely to me.

I want a “Duolingo for linguists”that doesn’t attempt to teach you useful everyday language but just speed runs you through the grammar and so on

Google translate gives “在这里，说得很多，但听到的却很少”, which has the suspicious property that the passive is marked with 得 in the first clause and 的 in the second. More credence to the theory that they are cognate with all Irish, specifically the Cork dialect, which is the oldest and purest form

Interesting, my brain didn’t consider that because that construction feels like an adjective rather than a verb, but it does seem to have more or less the same meaning!

I often find myself wanting to use the Irish impersonal aspect in Chinese, a.k.a. 说tar/听tear for "it is spoken"/"it is heard", so ”在这里说tar很多，而听tear很少“ “here, much is said, but little is heard". (While we're at it we could give Irish a nice Rechtschreibreform imported from Cyrl: "听tear" --> "听tьar")