Tatar instructions

A Tatar instruction-tuning dataset in the InstruCAT format, with an extra conversations column that holds the chat messages as a JSON string. All instructions are in Tatar.

column	description
id	<source>_<source id>, e.g. wikigec_31223 (the index in wikiedits_tt_gec.parquet)
instruction	task instruction (Tatar), one of several templates, picked deterministically from the id
context	input text
response	expected answer
category	task category
conversations	JSON string [{"role": "user", "content": instruction + "\n\n" + context}, {"role": "assistant", "content": response}]

category	train	test	source
gec	22,541	2,506	WikiGEC-TT — Tatar Wikipedia edit history 2003–2026, script=cyrl, edit_type=spelling; split by article (no page in both splits)
toxicity_classification	4,490	500	textdetox/multilingual_toxicity_dataset tt; stratified 90/10; answers «Әйе»/«Юк»
emotion_classification	1,193	1,000	brighter-dataset/BRIGHTER-emotion-categories tat; original splits (train+dev → train, test → test); labels ачу, җирәнү, курку, шатлык, кайгы, гаҗәпләнү / «хис юк»
detoxification	475	120	textdetox/multilingual_paradetox_test tt; random 80/20
total	28,699	4,126

Processing: whitespace normalization, empty rows and rows over 4,000 characters removed, deduplication on (context, response), and a cross-source leakage guard so that no train context also occurs in test.

import json
from datasets import load_dataset
ds = load_dataset("gaydmi/instruct")
messages = json.loads(ds["train"][0]["conversations"])

Notes:

• gec is silver data mined from human Wikipedia edits. The spelling subset is the cleanest part.
• The detoxification pairs come from the official TextDetox ParaDetox test set. Models trained on its train part should not be evaluated on that benchmark.
• Licenses follow the sources: Wikipedia is CC BY-SA 4.0; check the TextDetox and BRIGHTER cards for theirs.