averoo/sumi

SuperMinor — assistant-style instruction prompts in 15 minority languages (RU pivot)

A fully parallel set of 124 assistant-style instruction prompts (summarize, extract key points, quiz questions, style transfer, translation requests, etc.) translated from Russian into 15 low-resource languages, most of them minority languages of Russia. Every prompt has a translation in every language — a perfect N-way parallel table.

The prompts imitate how a user talks to a virtual assistant. Together with per-language answers they can be used to teach instruction-following models to converse in languages that today's assistants barely speak.

About the SuperMinor project

SuperMinor is a project for the development of minor languages, aimed at collecting and translating data for training language models in instruction format. Virtual assistants mostly work well only with popular languages — and not even all of those. For small languages like Yakut or Chuvash the situation is far worse.

For the project, a small portal was built where native speakers of minor languages could, without any extra hassle, translate and edit specially selected short texts in the form of instructions. The instructions are sets of phrases imitating a conversation with an assistant. By training on such texts, a model learns to communicate in a language new to it, which makes it possible to build all kinds of generative content and neural-network applications in Bashkir, Tatar, the Mari languages, the Ossetian dialects, and the other languages for which the data was collected.

This dataset is the result of the project: native-speaker translations, validated with the Claude Fable 5 model, which also produced the translations (fable rows) for the prompts no volunteer had covered.

Languages

code	language	human	fable
ru	Russian (pivot)	—	—
alt	Southern Altai	124	0
av	Avar	2	122
ba	Bashkir	124	0
be	Belarusian	5	119
cv	Chuvash	124	0
dar	Dargwa	124	0
dig	Ossetian (Digor)	0	124
kjh	Khakas	25	99
kk	Kazakh	124	0
kv	Komi-Zyrian	124	0
mhr	Meadow Mari	124	0
os	Ossetian (Iron)	0	124
sah	Yakut	124	0
tt	Tatar	124	0
udm	Udmurt	124	0

Codes are ISO 639-1 where available, ISO 639-2/3 otherwise. os is Iron Ossetian, the literary standard; Digor Ossetian has no dedicated ISO code, so dig follows the convention of the companion Little Prince multiparallel corpus of low resource languages of Russia.

Data

data/prompts.parquet contains one flat row per prompt. Each language is a text column and carries two metadata columns prefixed with its code:

{
  "ru": "Выпиши десять ключевых слов, описывающих текст:",
  "ba": "Тексты тасуирлаған ун төп һүҙҙе яҙып ал:",
  "ba_source": "human",
  "ba_edited": false
}

• <code> — the prompt in that language;
• <code>_source — human (translated by a native-speaker volunteer on the SuperMinor portal) or fable (no volunteer translation existed; translated by Claude Fable 5);
• <code>_edited — true if the text was touched during the validation pass (typo / consistency / mistranslation fixes for human rows; always true for fable rows).

When pristine native-speaker text is required, filter on <code>_source == "human".

Validation

Every human translation was reviewed and every gap filled in a per-language verification pass (Claude Fable 5), using the project's Little Prince multiparallel corpus and its per-language dictionaries as orthographic and lexical reference. The pass enforced: faithfulness to the Russian prompt, one orthographic convention per language (single palochka form, Cyrillic-only letters where standard, one че/nasal letter for Khakas, degemination per native majority for Bashkir), an exact XXX placeholder, digit and trailing-colon parity with the pivot, and terminology consistency across each language's prompt set.

Usage

from datasets import load_dataset

prompts = load_dataset("averoo/sumi", "prompts")["train"]
row = prompts[0]
print(row["ru"], "->", row["sah"])

Licensing and intended use

Volunteer translations were contributed to the SuperMinor project for developing language technology for minority languages. Published for language preservation, education, and research; attribution appreciated.