Amharic
Collection
4 items • Updated
amh-post-training-v2
Version 2 of SaarAI/amh-post-training, with the same row format.
v1 was used to fine-tune SaarAI/AfriqueQwen3.5-4B-amh-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.
802,189 train / 5,565 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 |
|---|---|
| SIB -38 (no SIB-style data in v1) | sib200_bench: SIB-200 amh_Ethi train with English labels, label-balanced, shuffled label lists |
| AfriXNLI -23, same option picked 96% of the time | multinli_en: label-balanced English MultiNLI sample (AfriXNLI has no train split) |
| AfriSenti -16; v1 sentiment is 53% neutral and mostly Amharic labels | afrisenti_bench: AfriSenti train with English labels, balanced; v1 sentiment labels rebalanced |
| Amharic->English MT -2.9 chrF; Walia/habtew prompts only | bare Amharic: ...\nEnglish: MT rows (*_bare); SMOL am<->en sentence and document pairs added |
| MasakhaNEWS +20, intent +4 (kept) | masakhanews_bench, injongo_intent_bench: English-label variants, balanced; v1 news rebalanced |
| MMLU / MGSM flat | Aya CoT / QA subsets raised from 10k to up to 30k (afriinstruct_*) |
| code chat (24% of v1) and afridocmt (13%) dominate the mix | code170k_amh capped at 80k rows, afridocmt_amh at 50k |
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, MasakhaNEWS).
- 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 (amh / en = template language, native = source already an
instruction set), instruction, input, output, messages.
Tasks
| task | rows |
|---|---|
| chat | 152,868 |
| question_answering | 105,436 |
| code_chat | 80,000 |
| instruction_following | 63,397 |
| amharic_mt amh-eng | 63,276 |
| amharic_mt eng-amh | 63,136 |
| document_translation | 49,558 |
| generation | 30,866 |
| natural_language_inference | 21,000 |
| amharic_mezmur_completion | 17,888 |
| sentiment_analysis | 16,317 |
| amharic_title_generation | 14,580 |
| news_topic_classification | 10,620 |
| amharic_spellcheck | 10,508 |
| xlsum_summerization | 10,357 |
| xlsum_reverse_summerization | 10,347 |
| masakhaner | 10,244 |
| dialogue | 10,000 |
| text_simplification | 9,998 |
| afrisent | 7,320 |
| summarization | 7,065 |
| masakhanews | 6,925 |
| event_linking | 6,586 |
| amharic_mezmur_generation | 5,039 |
| intent_classification | 4,321 |
| amharic_poem_completion | 3,776 |
| topic_classification | 3,491 |
| amharicqa | 2,317 |
| amharic_story_generation | 1,974 |
| roleplay | 1,923 |
| story_generation | 1,713 |
| story_choice | 1,713 |
| amharic_poem_generation | 1,126 |
| translation | 1,035 |
| paraphrase_identification | 635 |
| amharic_zefen_completion | 264 |
| amharic_zefen_generation | 116 |
| text_classification | 19 |
Sources
New in v2: sib200_bench (Davlan/sib200 train), injongo_intent_bench (masakhane/InjongoIntent train),
multinli_en (nyu-mll/multi_nli train), afrisenti_bench (masakhane/afrisenti train), masakhanews_bench (masakhane/masakhanews train), smol_*_amh (google/smol am/en), *_bare (bare-prompt MT made from the named v1 source) and more
afriinstruct_* rows (CohereLabs/aya_collection_language_split). See the v1 card for all other sources.
| source | rows |
|---|---|
| code170k_amh | 80,000 |
| addisai_alpaca | 49,809 |
| addisai_finetome | 49,640 |
| afridocmt_amh | 49,558 |
| amharic_alpaca | 46,546 |
| habtew_mt | 39,505 |
| habtew_mt_bare | 35,028 |
| afriinstruct_HotpotQA_T | 32,662 |
| afriinstruct_NQ_Open_T | 31,424 |
| afriinstruct_Flan_CoT_submix_T | 30,676 |
| kakugo_amh | 29,044 |
| multinli_en | 21,000 |
| walia_amharic_mezmur_completion | 17,888 |
| afriinstruct_PIQA_T | 16,161 |
| amharic_dolly | 14,991 |
| walia_amharic_title_generation | 14,580 |
| afriinstruct_Mintaka_inst_T | 13,914 |
| walia_amharic_mt amh-eng | 10,858 |
| walia_amharic_mt eng-amh | 10,858 |
| walia_amharic_mt eng-amh_bare | 10,779 |
| walia_amharic_spellcheck | 10,508 |
| walia_amharic_mt amh-eng_bare | 10,425 |
| walia_xlsum_summerization | 10,357 |
| walia_xlsum_reverse_summerization | 10,347 |
| walia_masakhaner | 10,244 |
| afriinstruct_SODA_inst_T | 10,000 |
| afriinstruct_Wiki_split_inst_T | 9,998 |
| afriinstruct_Adversarial_QA_T | 9,930 |
| afrisenti_bench | 8,849 |
| afriinstruct_AfriSenti | 7,341 |
| walia_afrisent | 7,320 |
| walia_masakhanews | 6,925 |
| daily_dialog_amh | 6,834 |
| afriinstruct_Xlel_wd_inst_T | 6,586 |
| afriinstruct_MasakhaNEWS | 6,020 |
| afriinstruct_XL_Sum | 5,557 |
| walia_amharic_mezmur_generation | 5,039 |
| muller_conversations | 5,000 |
| masakhanews_bench | 4,600 |
| african_ultrachat_amh | 4,529 |
| injongo_intent_bench | 4,321 |
| walia_amharic_poem_completion | 3,776 |
| sib200_bench | 3,491 |
| amharic_storyqa | 3,426 |
| multiturn_combined_amh | 3,375 |
| smol_sent_amh | 3,199 |
| smol_sent_amh_bare | 3,106 |
| tadesse_amharic_qa | 2,617 |
| smol_doc_amh | 2,539 |
| walia_amharicqa | 2,317 |
| walia_amharic_story_generation | 1,974 |
| roleplay_amh | 1,923 |
| afriinstruct_CNN_Daily_Mail_T | 1,228 |
| aya_human_amh | 1,180 |
| walia_amharic_poem_generation | 1,126 |
| afriinstruct_MAFAND | 1,035 |
| amharic_maternal_health | 1,011 |
| wikiconversations_amh | 1,009 |
| amharic_srh_qa | 991 |
| addisgpt_amh | 680 |
| afriinstruct_PAWS_Wiki_T | 635 |
| afriinstruct_Flan_GEM_wiki_lingua_T | 280 |
| walia_amharic_zefen_completion | 264 |
| afriinstruct_MLQA_en_T | 250 |
| afriinstruct_Dolly_v2_T | 140 |
| afriinstruct_AfriSenti_inst | 127 |
| walia_amharic_zefen_generation | 116 |
| smol_doc_amh_bare | 115 |
| afriinstruct_Flan_Coqa_T | 70 |
| afriinstruct_Flan_lambada_T | 50 |
| afriinstruct_WIKI_QA_T | 34 |
| afriinstruct_MasakhaNEWS_inst | 19 |
Removed / filtered rows
| reason | rows |
|---|---|
| code170k_amh: down-weighted to 80,000 (v2 cap) | 95,045 |
| afridocmt_amh: down-weighted to 50,000 (v2 cap) | 40,360 |
| walia_afrisent: label 'ገለልተኛ' rebalanced to median 2,932 | 3,118 |
| afriinstruct_AfriSenti: label 'neutral' rebalanced to median 2,595 | 2,539 |
| decontamination: overlaps AfroBench afrisenti (walia_afrisent) | 1,058 |
| habtew_mt: bare-MT skipped, wrong script | 581 |
| decontamination: overlaps AfroBench masakhanews (walia_masakhanews) | 563 |
| decontamination: overlaps AfroBench masakhanews (walia_xlsum_reverse_summerization) | 322 |
| decontamination: overlaps AfroBench masakhanews (walia_amharic_spellcheck) | 320 |
| decontamination: overlaps AfroBench masakhanews (walia_xlsum_summerization) | 312 |
| decontamination: overlaps AfroBench masakhanews (walia_amharic_title_generation) | 177 |
| decontamination: overlaps AfroBench masakhanews (afriinstruct_XL_Sum) | 172 |
| decontamination: overlaps AfroBench sib (walia_amharic_mt amh-eng) | 137 |
| decontamination: overlaps AfroBench sib (walia_amharic_mt eng-amh) | 137 |
| decontamination: overlaps AfroBench sib (walia_amharic_mt eng-amh_bare) | 137 |
| decontamination: overlaps AfroBench sib (walia_amharic_mt amh-eng_bare) | 137 |
| walia_masakhanews: label 'ይህ ዜና ፖለቲካ ነው' rebalanced to median 2,009 | 134 |
| decontamination: overlaps AfroBench afrisenti (afrisenti_bench) | 134 |
| decontamination: overlaps AfroBench masakhanews (masakhanews_bench) | 128 |
| decontamination: overlaps AfroBench afrisenti (afriinstruct_AfriSenti) | 119 |
| afriinstruct_Flan_CoT_submix_T (v2 extra): not target language (GlotLID) | 116 |
| afriinstruct_MasakhaNEWS: label 'politics' rebalanced to median 1,635 | 115 |
| decontamination: overlaps AfroBench masakhanews (afriinstruct_MasakhaNEWS) | 90 |
| afriinstruct_Adversarial_QA_T (v2 extra): empty / None / decoding collapse | 68 |
| decontamination: overlaps AfroBench afrisenti (afriinstruct_AfriSenti_inst) | 63 |
| afriinstruct_Flan_CoT_submix_T (v2 extra): empty / None / decoding collapse | 46 |
| walia_amharic_mt amh-eng: bare-MT skipped, wrong script | 44 |
| smol_sent_amh: bad pair (length ratio / too short / identical) | 40 |
| afriinstruct_PIQA_T (v2 extra): empty / None / decoding collapse | 38 |
| walia_amharic_mt eng-amh: bare-MT skipped, wrong script | 38 |
| decontamination: overlaps AfroBench afrisenti (walia_xlsum_summerization) | 29 |
| decontamination: overlaps AfroBench afrisenti (walia_xlsum_reverse_summerization) | 26 |
| decontamination: overlaps AfroBench afrisenti (walia_amharic_spellcheck) | 24 |
| afriinstruct_HotpotQA_T (v2 extra): not target language (GlotLID) | 20 |
| decontamination: overlaps AfroBench afrisenti (afriinstruct_XL_Sum) | 17 |
| decontamination: overlaps AfroBench belebele_qa (sib200_bench) | 16 |
| decontamination: overlaps AfroBench afrisenti (walia_amharic_title_generation) | 14 |
| decontamination: overlaps AfroBench masakhanews (afriinstruct_MasakhaNEWS_inst) | 11 |
| decontamination: overlaps AfroBench afrisenti (aya_human_amh) | 9 |
| afriinstruct_NQ_Open_T (v2 extra): not target language (GlotLID) | 8 |
| afriinstruct_PIQA_T (v2 extra): not target language (GlotLID) | 7 |
| decontamination: overlaps AfroBench afrisenti (walia_masakhanews) | 6 |
| decontamination: overlaps AfroBench injongointent (injongo_intent_bench) | 6 |
| decontamination: overlaps AfroBench masakhanews (aya_human_amh) | 3 |
| decontamination: overlaps AfroBench sib (sib200_bench) | 3 |
| decontamination: overlaps AfroBench belebele_qa (afriinstruct_PIQA_T) | 2 |
| afriinstruct_HotpotQA_T (v2 extra): empty / None / decoding collapse | 1 |
| afriinstruct_Mintaka_inst_T (v2 extra): not target language (GlotLID) | 1 |
| decontamination: overlaps AfroBench afrimmlu (amharic_alpaca) | 1 |
| decontamination: overlaps AfroBench belebele_qa (walia_amharic_mt eng-amh) | 1 |
| decontamination: overlaps AfroBench afrimmlu (addisai_finetome) | 1 |
| decontamination: overlaps AfroBench afrisenti (kakugo_amh) | 1 |
| decontamination: overlaps AfroBench masakhanews (wikiconversations_amh) | 1 |
| decontamination: overlaps AfroBench afrisenti (walia_amharic_mt eng-amh) | 1 |
| decontamination: overlaps AfroBench belebele_qa (walia_amharic_mt amh-eng) | 1 |
| decontamination: overlaps AfroBench afrisenti (walia_amharic_mt amh-eng) | 1 |
| decontamination: overlaps AfroBench belebele_qa (walia_amharic_mt eng-amh_bare) | 1 |
| decontamination: overlaps AfroBench afrisenti (walia_amharic_mt eng-amh_bare) | 1 |
| decontamination: overlaps AfroBench afrisenti (walia_amharic_mt amh-eng_bare) | 1 |
| decontamination: overlaps AfroBench belebele_qa (walia_amharic_mt amh-eng_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
- 243