zul-post-training-v2
Version 2 of SaarAI/zul-post-training, with the same row format.
v1 was used to fine-tune SaarAI/AfriqueQwen3.5-4B-zul-sft and evaluated on AfroBench. v2 keeps v1
except where noted below and adds data aimed at the problems found in that evaluation. The whole set is meant for
training; v2 is not limited to a 100k sample.
858,985 train / 6,645 validation rows. Validation is a per-task 0.5% holdout (20-1000 rows per task), split by item: all prompt variants and both MT directions of one item are on the same side.
v2 changes
| problem seen in the v1 model | change in v2 |
|---|---|
| Zulu->English MT: model rewrote the input in Zulu (wrong language 13.5% -> 39.6%, chrF -7.7); all v1 MT prompts were full instructions | bare Zulu: ...\nEnglish:-style MT rows (*_bare), best corpora first (SMOL, govza, then Autshumato); SMOL zu->en (Zulu-original) pairs added |
| MasakhaPOS -17: 34% of outputs unparsable (word/TAG instead of tuples) | masakhapos_tuples: MasakhaPOS train answered as [('word', 'TAG'), ...] (3 prompt variants each) |
| Small, skewed classification sets (SIB 701, news 563 rows) | sib200_bench, injongo_intent_bench: English-label, shuffled-label, label-balanced variants; v1 classification labels rebalanced to the median label count |
| AfriXNLI: no NLI data in v1 | multinli_en: label-balanced English MultiNLI sample (AfriXNLI has no train split) |
| MMLU -2.4, MGSM -1.6 | Aya CoT / QA subsets raised from 10k to up to 30k (aya_*) |
| Autshumato (40% of v1) and code chat (18%) dominate the mix | autshumato_zul capped at 200k rows, code170k_zul at 80k |
Also:
- Exact duplicates are found on (task, instruction, input, output), so different prompt variants of one item are kept.
- Decontamination: rows that share a 10-word n-gram (or a whole short text of >= 6 words) with an AfroBench test/dev item are removed (SIB, AfriXNLI, InjongoIntent, AfriMGSM, AfriMMLU, NTREX, Belebele questions and answers, MasakhaPOS).
- SIB-200 and Belebele are both built from FLORES-200: about 90% of SIB train sentences appear inside Belebele passages. Belebele passages are therefore not a decontamination source (that would remove all SIB data), and Belebele gains after training on this set (or on v1) may be inflated.
- Benchmark-style prompts were written for this dataset in the style of AfroBench prompt_1. They are paraphrases,
not copies, and label lists are shuffled.
multinli_enrows are English.
Columns
id, source, task, instruction_lang (zul / en = template language, native = source already an
instruction set), instruction, input, output, messages.
Tasks
| task | rows |
|---|---|
| translation_zul_en | 143,217 |
| translation_en_zul | 143,139 |
| question_answering | 123,212 |
| chat | 88,622 |
| instruction_following | 83,267 |
| code_chat | 80,000 |
| generation | 51,109 |
| text_continuation | 39,995 |
| natural_language_inference | 21,000 |
| summarization | 19,884 |
| dialogue | 9,996 |
| text_simplification | 9,976 |
| paraphrase_identification | 9,928 |
| event_linking | 9,733 |
| intent_classification | 6,539 |
| named_entity_recognition | 5,844 |
| translation | 5,216 |
| topic_classification | 4,045 |
| pos_tagging | 2,995 |
| hate_speech_detection | 2,366 |
| title_generation | 1,791 |
| article_generation | 1,791 |
| aya_human | 1,652 |
| news_topic_classification | 313 |
Sources
New in v2: sib200_bench (Davlan/sib200 train), injongo_intent_bench (masakhane/InjongoIntent train),
multinli_en (nyu-mll/multi_nli train), masakhapos_tuples (masakhane/masakhapos train), smol_*_zu_orig (google/smol zu_en), *_bare (bare-prompt MT made from the named v1 source) and more
aya_* rows (CohereLabs/aya_collection_language_split). See the v1 card for all other sources.
| source | rows |
|---|---|
| autshumato_zul | 199,959 |
| code170k_zul | 80,000 |
| autshumato_zul_bare | 41,023 |
| alpaca_zul | 40,317 |
| fineweb2_zul | 39,995 |
| smoltalk_zul | 39,649 |
| aya_HotpotQA_T | 37,988 |
| aya_NQ_Open_T | 36,869 |
| aya_Flan_CoT_submix_T | 36,530 |
| kakugo_zul | 36,127 |
| self_instruct_zul | 28,878 |
| multinli_en | 21,000 |
| govza_cabinet | 17,396 |
| govza_cabinet_bare | 17,220 |
| aya_PIQA_T | 16,031 |
| dolly_zul | 14,072 |
| aya_Mintaka_inst_T | 13,855 |
| aya_SODA_inst_T | 9,996 |
| aya_Adversarial_QA_T | 9,982 |
| aya_Wiki_split_inst_T | 9,976 |
| aya_CNN_Daily_Mail_T | 9,949 |
| aya_Flan_GEM_wiki_lingua_T | 9,935 |
| aya_PAWS_Wiki_T | 9,928 |
| aya_Xlel_wd_inst_T | 9,733 |
| aya_Dolly_v2_T | 9,689 |
| aya_Flan_Coqa_T | 6,408 |
| masakhaner2 | 5,844 |
| inkuba_mafand | 5,216 |
| african_ultrachat_zul | 4,899 |
| vukuzenzele_zul | 4,404 |
| injongo_intent_bench | 4,305 |
| aya_Flan_lambada_T | 4,156 |
| wikipedia_zu | 3,582 |
| multiturn_combined_zul | 3,547 |
| sib200_bench | 3,510 |
| gsma_conversations_zul | 2,796 |
| afrihate | 2,366 |
| masakhapos_tuples | 2,244 |
| injongo_intent | 2,234 |
| smol_sent | 1,724 |
| smol_sent_bare | 1,723 |
| aya_Aya_Dataset | 1,652 |
| wikiconversations_zul | 1,604 |
| smol_sent_zu_orig | 1,573 |
| aya_WIKI_QA_T | 1,037 |
| masakhapos | 751 |
| aya_Joke_explaination_inst_T | 734 |
| smol_doc | 678 |
| smol_doc_zu_orig | 622 |
| aya_Flan_unified_QA_T | 536 |
| sib200 | 535 |
| za_isizulu_news | 313 |
| inkuba_afriqa | 296 |
| aya_AfriQA_inst | 210 |
| smol_doc_bare | 34 |
Removed / filtered rows
| reason | rows |
|---|---|
| autshumato_zul: down-weighted to 200,000 (v2 cap) | 190,256 |
| code170k_zul: down-weighted to 80,000 (v2 cap) | 96,492 |
| aya_NQ_Open_T (v2 extra): not target language (GlotLID) | 3,481 |
| aya_Flan_CoT_submix_T (v2 extra): not target language (GlotLID) | 567 |
| aya_HotpotQA_T (v2 extra): not target language (GlotLID) | 234 |
| afrihate: label 'okujwayelekile' rebalanced to median 586 | 195 |
| afrihate: label 'normal' rebalanced to median 586 | 176 |
| za_isizulu_news: label 'ubugebengu, umthetho nobulungiswa' rebalanced to median 40 | 118 |
| za_isizulu_news: label 'crime, law and justice' rebalanced to median 40 | 94 |
| aya_PIQA_T (v2 extra): not target language (GlotLID) | 71 |
| sib200: label 'science/technology' rebalanced to median 42 | 51 |
| sib200: label 'isayensi nobuchwepheshe' rebalanced to median 42 | 41 |
| sib200: label 'ezokuvakasha' rebalanced to median 42 | 38 |
| za_isizulu_news: label 'politics' rebalanced to median 40 | 20 |
| sib200: label 'ezombusazwe' rebalanced to median 42 | 18 |
| za_isizulu_news: label 'ezombusazwe' rebalanced to median 40 | 18 |
| sib200: label 'travel' rebalanced to median 42 | 16 |
| aya_Mintaka_inst_T (v2 extra): not target language (GlotLID) | 12 |
| decontamination: overlaps AfroBench injongointent (injongo_intent_bench) | 12 |
| aya_Adversarial_QA_T (v2 extra): not target language (GlotLID) | 6 |
| aya_PIQA_T (v2 extra): empty / None / decoding collapse | 6 |
| decontamination: overlaps AfroBench injongointent (injongo_intent) | 6 |
| decontamination: overlaps AfroBench masakhapos (masakhapos_tuples) | 6 |
| aya_Adversarial_QA_T (v2 extra): empty / None / decoding collapse | 5 |
| aya_Flan_CoT_submix_T (v2 extra): empty / None / decoding collapse | 5 |
| decontamination: overlaps AfroBench masakhapos (fineweb2_zul) | 5 |
| masakhapos_tuples: token contains a quote (tuple format would not parse) | 3 |
| sib200: label 'ezemidlalo' rebalanced to median 42 | 2 |
| decontamination: overlaps AfroBench masakhapos (masakhapos) | 2 |
| smol_sent_zu_orig: bad pair (length ratio / too short / identical) | 1 |
| decontamination: overlaps AfroBench ntrex (aya_CNN_Daily_Mail_T) | 1 |
| decontamination: overlaps AfroBench ntrex (smoltalk_zul) | 1 |
| decontamination: overlaps AfroBench masakhapos (kakugo_zul) | 1 |
Known limitations
- All v1 limitations still apply (machine-translated and synthetic sources; licenses differ per source; see the v1 card).
- Only exact duplicates and AfroBench n-gram overlaps are removed; near-duplicates may remain.
- The benchmark-style rows make AfroBench prompt_1 easier; also check other prompts and held-out tasks.
- Some sources are gated or private, so this repo is gated like v1.
- Downloads last month
- 763