kin-post-training

Kinyarwanda instruction / chat data compiled from public datasets in the style of Walia-LLM (Azime et al., 2024, arXiv:2402.08015): task datasets (translation, classification, tagging, NER, generation) are converted to instructions using many paraphrased templates, and existing instruction / chat sets are added as-is.

1,002,844 train / 3,743 validation rows. Validation is a random per-task holdout (0.5%, 20-1000 rows per task).

Columns

id, source, task, instruction_lang (kin or en = language of the template, native = source already an instruction/chat set), instruction, input, output, messages (chat form: list of role/content dicts, multi-turn for chat sources).

Tasks

task	rows
translation_kin_en	336,742
translation_en_kin	336,288
code_chat	176,678
chat	43,321
instruction_following	41,601
text_continuation	40,271
news_topic_classification	9,075
sentiment_analysis	6,536
article_generation	5,916
title_generation	5,916
intent_classification	2,240
topic_classification	800
pos_tagging	753
question_answering	255
named_entity_recognition	195

Sources

source key	dataset	note
du_gazette	DigitalUmuganda/NMT_Rwandan-Gazette_parallel_data_en_kin	human-validated en-kin legal gazette parallel text
du_health	DigitalUmuganda/NMT_Health_parallel_data_en_kin	human translations, health domain
rogerb_du	RogerB/kin_en_DigitalUmuganda	en-kin sentence pairs (Digital Umuganda)
kira_en_rw	Kira-Floris/cross-lingual-translation-en-kin-parquet	en-kin translation instructions (150k sample of 492k)
smalltalk_phrases	map-boy/kinyarwanda-dialogue-dataset	short small-talk phrase pairs
code170k_kin	michsethowusu/Code-170k-kinyarwanda	machine-translated code Q&A chat
alpaca_kin	saillab/alpaca-kinyarwanda-cleaned	machine-translated Alpaca
kakugo_kin	ptrdvn/kakugo-kin	synthetic LLM chat (reasoning blocks removed)
african_ultrachat	masakhane/african-ultrachat	multi-turn chat
mbaza_mono	mbazaNLP/kinyarwanda_monolingual_v01.0	monolingual text -> text continuation (40k sample)
wikipedia_rw	wikimedia/wikipedia (20231101.rw)	Wikipedia -> article/title/continuation tasks
kinnews	KoseiUemura/KinNewsClassification (kinnews_cleaned)	news topic classification (14 classes)
sib200	Davlan/sib200 (kin_Latn)	topic classification (7 classes)
injongo_intent	masakhane/InjongoIntent (kin)	intent classification (40 intents)
masakhapos	masakhane/masakhapos (kin)	UPOS tagging
wikiann	unimelb-nlp/wikiann (rw)	named entity recognition
aya_	CohereLabs/aya_collection_language_split (kinyarwanda)	AfriSenti sentiment + AfriQA (instructions in English)

Known limitations (please read)

• The Kinyarwanda instruction templates (templates.json) were drafted by an AI model and have not been reviewed by a native speaker.
• Quality varies by source: code170k_kin and alpaca_kin are machine-translated; kakugo_kin is synthetic; kira_en_rw mixes human and machine translations. Use the source column to filter or reweight.
• Excluded after inspection: Helsinki-NLP/opus-100 en-rw (noisy machine-translated software strings) and mbazaNLP/Kinyarwanda_English_parallel_dataset (English side is lower-cased Bible text with stopwords removed).
• Gated or unreadable sources were not included. Licenses differ per source; check each source dataset before commercial use.
• Only exact duplicates were removed (per task direction); there may be near-duplicates and overlap between sources. Test splits of source datasets were not used, but SIB-200 dev and WikiANN validation were, so do not evaluate on those.