kin-post-training-v2
Version 2 of SaarAI/kin-post-training, with the same row format.
v1 was used to fine-tune SaarAI/afriqueqwen3.5-4b-kin-full, but that model has not been evaluated on
AfroBench yet. The changes below fix the v1 data issues that caused AfroBench drops for the Zulu and Amharic
models trained the same way (same base model, same recipe, Walia-style data). The whole set is meant for
training; v2 is not limited to a 100k sample.
715,447 train / 4,482 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
| v1 data issue (and the effect seen for zul / amh) | change in v2 |
|---|---|
| All v1 MT prompts are full instructions (zul model then rewrote the input instead of translating: Zulu->English -7.7 chrF) | bare Kinyarwanda: ...\nEnglish:-style MT rows (*_bare), best corpora first (SMOL, health, Digital Umuganda, Kira, gazette); SMOL en<->rw sentence and document pairs added (not in v1) |
MasakhaPOS answers are word/TAG (zul model: 34% of outputs unparsable, POS -17) |
masakhapos_tuples: MasakhaPOS kin train answered as [('word', 'TAG'), ...]; the bare ' token (Kinyarwanda elisions) is written as ’ so the tuple format stays parsable |
| Half of the SIB rows use Kinyarwanda label names, and AfroBench scores the English label (amh model with no English-label SIB data: SIB -38) | sib200_bench, injongo_intent_bench: English-label, shuffled-label, label-balanced variants; v1 KinNews / SIB / intent labels rebalanced to the median label count |
Sentiment rows answer with a sentence (I would classify the given tweet as: positive.), each tweet twice (amh model: AfriSenti -16) |
afrisenti_bench: AfriSenti kin train answered with the bare English label, label-balanced |
| No NLI data (amh model picked one AfriXNLI option 96% of the time) | multinli_en: label-balanced English MultiNLI sample (AfriXNLI has no train split) |
| MT is 67% of v1 (legal gazette alone 32%), code chat 18% | du_gazette capped at 120k rows, du_health and kira_en_rw at 80k, code170k_kin at 80k |
| Chat is 4% of v1; the GSMA / multiturn / Wiki conversations were added outside the dataset in the SFT runs | gsma_conversations_kin (local hf_split train + validation; the test split is kept out for evaluation), multiturn_combined_kin, wikiconversations_kin added, filtered like v1 chat |
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, AfriSenti, 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 (kin / en = template language, native = source already an
instruction set), instruction, input, output, messages.
Tasks
| task | rows |
|---|---|
| translation_en_kin | 219,594 |
| translation_kin_en | 216,530 |
| code_chat | 79,999 |
| chat | 57,163 |
| instruction_following | 41,487 |
| text_continuation | 39,557 |
| natural_language_inference | 21,000 |
| sentiment_analysis | 14,301 |
| intent_classification | 6,557 |
| title_generation | 5,875 |
| article_generation | 5,850 |
| news_topic_classification | 4,825 |
| topic_classification | 3,854 |
| pos_tagging | 2,888 |
| question_answering | 254 |
| named_entity_recognition | 195 |
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), afrisenti_bench (masakhane/afrisenti train), smol_*_kin (google/smol en/rw), gsma_conversations_kin (GSMA pipeline, data/kinyarwanda_rw/hf_split), multiturn_combined_kin (israel/multiturn_combined), wikiconversations_kin (israel/WikiConversations), *_bare (bare-prompt MT made from the named v1 source).
See the v1 card for all other sources.
| source | rows |
|---|---|
| du_gazette | 119,937 |
| code170k_kin | 79,999 |
| du_health | 79,998 |
| kira_en_rw | 79,839 |
| rogerb_du | 78,406 |
| alpaca_kin | 41,487 |
| kakugo_kin | 38,124 |
| mbaza_mono | 34,813 |
| multinli_en | 21,000 |
| du_health_bare | 21,000 |
| rogerb_du_bare | 20,998 |
| wikipedia_rw | 16,469 |
| kira_en_rw_bare | 14,486 |
| smalltalk_phrases | 13,673 |
| gsma_conversations_kin | 9,615 |
| afrisenti_bench | 8,153 |
| aya_AfriSenti-inst | 6,148 |
| kinnews | 4,825 |
| african_ultrachat | 4,802 |
| injongo_intent_bench | 4,322 |
| multiturn_combined_kin | 3,849 |
| sib200_bench | 3,333 |
| smol_sent_kin | 3,285 |
| smol_sent_kin_bare | 3,155 |
| injongo_intent | 2,235 |
| masakhapos_tuples | 2,139 |
| smol_doc_kin | 1,309 |
| wikiconversations_kin | 773 |
| masakhapos | 749 |
| sib200_train | 521 |
| aya_AfriQA-inst | 254 |
| wikiann_train | 99 |
| wikiann_validation | 96 |
| smol_doc_kin_bare | 38 |
Removed / filtered rows
| reason | rows |
|---|---|
| du_gazette: down-weighted to 120,000 (v2 cap) | 204,512 |
| code170k_kin: down-weighted to 80,000 (v2 cap) | 96,678 |
| du_health: down-weighted to 80,000 (v2 cap) | 46,976 |
| kira_en_rw: down-weighted to 80,000 (v2 cap) | 46,158 |
| kinnews: label 'politiki' rebalanced to median 270 | 633 |
| kinnews: label 'politics' rebalanced to median 270 | 557 |
| decontamination: overlaps AfroBench belebele_qa (mbaza_mono) | 542 |
| decontamination: overlaps AfroBench afrisenti (afrisenti_bench) | 509 |
| kinnews: label 'imyidagaduro' rebalanced to median 270 | 493 |
| kinnews: label 'sport' rebalanced to median 270 | 481 |
| kinnews: label 'siporo' rebalanced to median 270 | 437 |
| kinnews: label 'entertainment' rebalanced to median 270 | 435 |
| decontamination: overlaps AfroBench afrisenti (aya_AfriSenti-inst) | 388 |
| kinnews: label 'ubukungu' rebalanced to median 270 | 287 |
| decontamination: overlaps AfroBench belebele_qa (kakugo_kin) | 284 |
| kinnews: label 'health' rebalanced to median 270 | 283 |
| kinnews: label 'ubuzima' rebalanced to median 270 | 281 |
| kinnews: label 'economy' rebalanced to median 270 | 233 |
| decontamination: overlaps AfroBench belebele_qa (sib200_bench) | 166 |
| decontamination: overlaps AfroBench belebele_qa (wikipedia_rw) | 163 |
| decontamination: overlaps AfroBench belebele_qa (kira_en_rw) | 123 |
| decontamination: overlaps AfroBench belebele_qa (alpaca_kin) | 113 |
| decontamination: overlaps AfroBench masakhapos (mbaza_mono) | 102 |
| decontamination: overlaps AfroBench belebele_qa (african_ultrachat) | 91 |
| decontamination: overlaps AfroBench sib (sib200_dev) | 89 |
| kinnews: label 'imibanire' rebalanced to median 270 | 60 |
| sib200_train: label 'science/technology' rebalanced to median 44 | 51 |
| decontamination: overlaps AfroBench belebele_qa (multiturn_combined_kin) | 44 |
| kinnews: label 'relationship' rebalanced to median 270 | 40 |
| sib200_train: label 'siyansi n'ikoranabuhanga' rebalanced to median 44 | 37 |
| kinnews: label 'religion' rebalanced to median 270 | 30 |
| decontamination: overlaps AfroBench belebele_qa (gsma_conversations_kin) | 28 |
| decontamination: overlaps AfroBench belebele_qa (sib200_train) | 27 |
| decontamination: overlaps AfroBench belebele_qa (du_gazette) | 26 |
| sib200_train: label 'ingendo' rebalanced to median 44 | 25 |
| sib200_train: label 'travel' rebalanced to median 44 | 25 |
| decontamination: overlaps AfroBench belebele_qa (wikiconversations_kin) | 21 |
| decontamination: overlaps AfroBench belebele_qa (kira_en_rw_bare) | 20 |
| decontamination: overlaps AfroBench masakhapos (kakugo_kin) | 16 |
| decontamination: overlaps AfroBench masakhapos (du_gazette) | 16 |
| decontamination: overlaps AfroBench afrisenti (mbaza_mono) | 11 |
| decontamination: overlaps AfroBench masakhapos (kira_en_rw) | 11 |
| decontamination: overlaps AfroBench sib (sib200_bench) | 10 |
| decontamination: overlaps AfroBench injongointent (injongo_intent_bench) | 10 |
| sib200_train: label 'politiki' rebalanced to median 44 | 8 |
| masakhapos_tuples: token contains a quote (tuple format would not parse) | 8 |
| decontamination: overlaps AfroBench belebele_qa (smol_doc_kin) | 8 |
| sib200_train: label 'politics' rebalanced to median 44 | 6 |
| wikiconversations_kin: not Kinyarwanda (GlotLID) | 6 |
| decontamination: overlaps AfroBench belebele_qa (rogerb_du) | 6 |
| decontamination: overlaps AfroBench belebele_qa (masakhapos_tuples) | 6 |
| decontamination: overlaps AfroBench masakhapos (masakhapos_tuples) | 6 |
| decontamination: overlaps AfroBench afrisenti (gsma_conversations_kin) | 6 |
| sib200_dev: label 'siyansi n'ikoranabuhanga' rebalanced to median 10 | 5 |
| decontamination: overlaps AfroBench injongointent (injongo_intent) | 5 |
| decontamination: overlaps AfroBench belebele_qa (sib200_dev) | 5 |
| multiturn_combined_kin: decoding collapse | 4 |
| decontamination: overlaps AfroBench masakhapos (multiturn_combined_kin) | 4 |
| multiturn_combined_kin: not Kinyarwanda (GlotLID) | 3 |
| decontamination: overlaps AfroBench masakhapos (wikipedia_rw) | 3 |
| decontamination: overlaps AfroBench afrisenti (kakugo_kin) | 3 |
| smol_sent_kin: bad pair (length ratio / too short / identical) | 2 |
| decontamination: overlaps AfroBench belebele_qa (masakhapos) | 2 |
| decontamination: overlaps AfroBench masakhapos (masakhapos) | 2 |
| decontamination: overlaps AfroBench belebele_qa (rogerb_du_bare) | 2 |
| decontamination: overlaps AfroBench belebele_qa (aya_AfriQA-inst) | 1 |
| decontamination: overlaps AfroBench ntrex (african_ultrachat) | 1 |
| decontamination: overlaps AfroBench sib (sib200_train) | 1 |
| decontamination: overlaps AfroBench belebele_qa (code170k_kin) | 1 |
| decontamination: overlaps AfroBench afrimmlu (alpaca_kin) | 1 |
| decontamination: overlaps AfroBench afrimmlu (kira_en_rw) | 1 |
| decontamination: overlaps AfroBench masakhapos (kira_en_rw_bare) | 1 |
| decontamination: overlaps AfroBench afrimmlu (kira_en_rw_bare) | 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
- 420