zul-post-training

isiZulu instruction / chat data compiled from public datasets in the style of Walia-LLM (Azime et al., 2024, arXiv:2402.08015), in the same row format as SaarAI/kin-post-training and SaarAI/amh-post-training: task datasets (translation, classification, tagging, NER, generation) are converted to instructions using paraphrased templates (templates.json), and existing instruction / chat sets are added as-is.

961,834 train / 4,859 validation rows. Validation is a random per-task holdout (0.5%, 20-1000 rows per task).

Columns

id, source, task, instruction_lang (zul 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; for multi-turn rows instruction / output are the first user / assistant turn).

Tasks

task	rows
translation_zul_en	207,229
translation_en_zul	207,229
code_chat	176,492
chat	88,624
instruction_following	83,267
question_answering	56,641
text_continuation	40,000
generation	24,405
summarization	19,885
dialogue	9,996
text_simplification	9,976
paraphrase_identification	9,928
event_linking	9,733
named_entity_recognition	5,844
translation	5,216
hate_speech_detection	2,737
intent_classification	2,240
title_generation	1,791
article_generation	1,791
aya_human	1,652
pos_tagging	753
topic_classification	701
news_topic_classification	563

Sources

source key	dataset	note
autshumato_zul	Mehgoss/sa-languages-translation (zu-en, autshumato)	en-zu government / health / agriculture sentence pairs
vukuzenzele_zul	dsfsi/vukuzenzele-sentence-aligned (eng-zul)	government-newspaper pairs (lower-cased); alignment score >= 0.8
govza_cabinet	dsfsi/govza-sa-cabinet-statements-sentence-aligned (eng-zul)	cabinet statements; alignment score >= 0.8
smol_sent, smol_doc	google/smol (smolsent / smoldoc en_zu)	professional en->zu sentence and document translations
sib200	Davlan/sib200 (zul_Latn)	topic classification (7 classes)
injongo_intent	masakhane/InjongoIntent (zul)	intent classification (40 intents)
za_isizulu_news	dsfsi/za-isizulu-siswati-news	Isolezwe headline topic classification (5 classes)
afrihate	afrihate/afrihate (zul)	hate / abuse detection (3 classes)
masakhapos	masakhane/masakhapos (zul)	UPOS tagging
masakhaner2	MasakhaNER 2.0 (zul train, GitHub)	named entity recognition (PER / LOC / ORG)
wikipedia_zu	wikimedia/wikipedia (20231101.zu)	Wikipedia -> article / title generation (>= 50 words)
fineweb2_zul	HuggingFaceFW/fineweb-2 (zul_Latn)	web text -> text continuation (40k sample)
gsma_conversations_zul	israel/zulu (train + validation)	grounded multi-agent conversations (GSMA pipeline)
multiturn_combined_zul	israel/multiturn_combined (zulu)	multi-turn conversations
wikiconversations_zul	israel/WikiConversations (isizulu)	multi-turn conversations grounded in Wikipedia
african_ultrachat_zul	masakhane/african-ultrachat (Zulu train)	multi-turn chat
kakugo_zul	ptrdvn/kakugo-zul	synthetic chat; <think> blocks removed
smoltalk_zul	2b2e/smoltalk-zulu-mtl	machine-translated SmolTalk (system turns dropped)
self_instruct_zul	2b2e/zulu-self_instruct-syn	synthetic self-instruct
code170k_zul	michsethowusu/Code-170k-zulu	machine-translated code Q&A chat
alpaca_zul	saillab/alpaca-zulu-cleaned	machine-translated Alpaca
dolly_zul	saillab/taco-datasets (Dolly-15k Zulu)	machine-translated Dolly
aya_*	CohereLabs/aya_collection_language_split (zulu train)	Aya mix (translated); each sub-dataset capped at 10k
inkuba_mafand, inkuba_afriqa	lelapa/Inkuba-instruct (isizulu_train)	MAFAND news MT and AfriQA, Inkuba's own isiZulu instructions

Row counts per source key:

source	rows
autshumato_zul	390,256
code170k_zul	176,492
alpaca_zul	40,317
fineweb2_zul	40,000
smoltalk_zul	39,650
kakugo_zul	36,128
self_instruct_zul	28,878
govza_cabinet	17,396
dolly_zul	14,072
aya_SODA_inst_T	9,996
aya_Wiki_split_inst_T	9,976
aya_Adversarial_QA_T	9,967
aya_CNN_Daily_Mail_T	9,950
aya_Flan_GEM_wiki_lingua_T	9,935
aya_PAWS_Wiki_T	9,928
aya_PIQA_T	9,926
aya_Mintaka_inst_T	9,911
aya_Flan_CoT_submix_T	9,826
aya_HotpotQA_T	9,801
aya_Xlel_wd_inst_T	9,733
aya_Dolly_v2_T	9,689
aya_NQ_Open_T	8,549
aya_Flan_Coqa_T	6,408
masakhaner2	5,844
inkuba_mafand	5,216
african_ultrachat_zul	4,899
vukuzenzele_zul	4,404
aya_Flan_lambada_T	4,156
wikipedia_zu	3,582
multiturn_combined_zul	3,547
gsma_conversations_zul	2,796
afrihate	2,737
injongo_intent	2,240
smol_sent	1,724
aya_Aya_Dataset	1,652
wikiconversations_zul	1,604
aya_WIKI_QA_T	1,037
masakhapos	753
aya_Joke_explaination_inst_T	734
sib200	701
smol_doc	678
za_isizulu_news	563
aya_Flan_unified_QA_T	536
inkuba_afriqa	296
aya_AfriQA_inst	210

Filtering

• Chat / instruction sources: system turns dropped, <think> blocks removed, conversations with an empty or literal "None" turn dropped, decoding-collapse (a word repeated 8+ times in a row, or very low compressibility) dropped, and GlotLID (cis-lmu/glotlid) must give zul_Latn >= 0.5 on the assistant text (code chat excluded).
• Translation pairs: zu/en character-length ratio 0.5-3.0, at least 2 words per side, identical pairs dropped.
• 14,978 exact duplicates were removed (same task + input/instruction + output; full conversation for chat).

reason	rows removed
aya_SODA_inst_T: over cap 10000	1,181,582
aya_Wiki_split_inst_T: over cap 10000	979,944
aya_Xlel_wd_inst_T: over cap 10000	513,112
aya_HotpotQA_T: over cap 10000	345,476
aya_NQ_Open_T: over cap 10000	165,850
aya_CNN_Daily_Mail_T: over cap 10000	90,000
aya_Flan_CoT_submix_T: over cap 10000	81,910
aya_PAWS_Wiki_T: over cap 10000	39,401
aya_Flan_GEM_wiki_lingua_T: over cap 10000	17,147
wikipedia_zu: < 50 words	9,770
aya_PIQA_T: over cap 10000	6,113
aya_Dolly_v2_T: over cap 10000	4,808
kakugo_zul: not isiZulu (GlotLID)	4,519
aya_Mintaka_inst_T: over cap 10000	4,000
self_instruct_zul: not isiZulu (GlotLID)	3,959
autshumato_zul: bad pair (length ratio / too short / identical)	3,509
govza_cabinet: alignment score < 0.8	3,493
smoltalk_zul: decoding collapse	2,728
smoltalk_zul: not isiZulu (GlotLID)	2,004
aya_NQ_Open_T: not isiZulu (GlotLID)	1,438
alpaca_zul: not isiZulu (GlotLID)	1,266
dolly_zul: not isiZulu (GlotLID)	918
kakugo_zul: malformed / empty / None turn	735
vukuzenzele_zul: alignment score < 0.8	421
aya_Dolly_v2_T: not isiZulu (GlotLID)	296
code170k_zul: malformed / empty / None turn	263
aya_Xlel_wd_inst_T: not isiZulu (GlotLID)	253
code170k_zul: decoding collapse	233
inkuba_afriqa: empty instruction / target	187
aya_Flan_CoT_submix_T: not isiZulu (GlotLID)	172
govza_cabinet: bad pair (length ratio / too short / identical)	134
aya_HotpotQA_T: not isiZulu (GlotLID)	126
aya_Flan_lambada_T: not isiZulu (GlotLID)	123
multiturn_combined_zul: not isiZulu (GlotLID)	115
aya_PIQA_T: not isiZulu (GlotLID)	67
self_instruct_zul: decoding collapse	64
aya_Flan_GEM_wiki_lingua_T: not isiZulu (GlotLID)	55
wikiconversations_zul: not isiZulu (GlotLID)	51
aya_CNN_Daily_Mail_T: not isiZulu (GlotLID)	34
aya_AfriQA_inst: not isiZulu (GlotLID)	29
aya_Wiki_split_inst_T: not isiZulu (GlotLID)	24
vukuzenzele_zul: bad pair (length ratio / too short / identical)	23
gsma_conversations_zul: not isiZulu (GlotLID)	20
aya_Adversarial_QA_T: not isiZulu (GlotLID)	18
aya_Aya_Dataset: not isiZulu (GlotLID)	11
alpaca_zul: decoding collapse	10
aya_Flan_GEM_wiki_lingua_T: decoding collapse	10
multiturn_combined_zul: decoding collapse	9
aya_Dolly_v2_T: decoding collapse	9
aya_Mintaka_inst_T: not isiZulu (GlotLID)	8
aya_PAWS_Wiki_T: not isiZulu (GlotLID)	8
aya_Xlel_wd_inst_T: decoding collapse	7
aya_Xlel_wd_inst_T: malformed / empty / None turn	7
dolly_zul: decoding collapse	5
aya_Adversarial_QA_T: decoding collapse	5
aya_Flan_unified_QA_T: not isiZulu (GlotLID)	4
aya_PIQA_T: decoding collapse	4
aya_SODA_inst_T: not isiZulu (GlotLID)	4
kakugo_zul: decoding collapse	3
aya_Flan_CoT_submix_T: decoding collapse	2
aya_WIKI_QA_T: not isiZulu (GlotLID)	2
smol_sent: bad pair (length ratio / too short / identical)	1
wikiconversations_zul: malformed / empty / None turn	1
african_ultrachat_zul: malformed / empty / None turn	1
aya_Flan_Coqa_T: decoding collapse	1

Known limitations (please read)

• The isiZulu instruction templates (templates.json) and translated label names were drafted by an AI model and have not been reviewed by a native speaker.
• Quality varies by source: code170k_zul, alpaca_zul, dolly_zul, smoltalk_zul and the aya_* (T) sets are machine-translated; kakugo_zul and self_instruct_zul are synthetic. Use the source column to filter or reweight.
• vukuzenzele_zul and inkuba_mafand text is lower-cased in the source. translation_zul_en outputs are English.
• Excluded after inspection: Helsinki-NLP/opus-100 en-zu (software UI strings), michsethowusu/english-zulu_sentence-pairs (many misaligned / garbage pairs), the 11.3M wmt22 rows of lelapa/Inkuba-instruct (noisy mined pairs, mostly zu->xx), and FLORES-derived data (benchmark). Benchmarks (AfriXNLI, AfriMMLU, AfriMGSM, Belebele, Uhura) were not used.
• Only train splits were used (plus israel/zulu validation); test splits of all sources (incl. israel/zulu test, SIB-200, InjongoIntent, AfriHate, MasakhaPOS/NER) remain usable for evaluation.
• Only exact duplicates were removed; near-duplicates and overlap between sources (e.g. Wikipedia vs. FineWeb-2, Aya vs. Inkuba) may remain. Licenses differ per source (Alpaca / Dolly / Aya derived data carry upstream terms; some sources have no license); check each before commercial use.