amh-post-training

Amharic post-training data: the Walia-LLM instruction dataset plus multi-turn conversations and other native instruction sets, in the same row format as SaarAI/kin-post-training.

714,181 train / 3,747 validation rows.

Columns

id, source, task, instruction_lang (native for all rows: instructions are already Amharic), instruction, input, output, messages (chat form; multi-turn for the conversation sources).

Tasks

task	rows
code_chat	175,045
chat	152,871
document_translation	90,360
instruction_following	63,410
amharic_mt amh-eng	30,825
amharic_mt eng-amh	30,676
amharic_mezmur_completion	17,888
amharic_title_generation	14,771
afrisent	11,496
amharic_spellcheck	10,852
xlsum_summerization	10,698
xlsum_reverse_summerization	10,695
masakhaner	10,244
sentiment_analysis	10,189
dialogue	10,000
text_simplification	9,998
masakhanews	7,628
question_answering	7,450
summarization	7,254
event_linking	6,586
news_topic_classification	6,225
amharic_mezmur_generation	5,039
amharic_poem_completion	3,776
amharicqa	2,317
amharic_story_generation	1,974
roleplay	1,923
story_choice	1,713
story_generation	1,713
amharic_poem_generation	1,126
generation	1,106
translation	1,035
paraphrase_identification	635
amharic_zefen_completion	264
amharic_zefen_generation	116
text_classification	30

Sources

source key	dataset	note
walia_* (17 tasks)	EthioNLP/Amharic_Instruction_dataset (train + validation splits)	Walia-LLM SFT data: native Amharic instructions for sentiment, NER, news classification, MT, summarization, spell-check, QA, title/poem/lyrics/story generation
multiturn_combined_amh	israel/multiturn_combined (amharic)	6-turn conversations; only the conversation column is used
wikiconversations_amh	israel/WikiConversations (amharic)	6-turn conversations; only the conversation column is used
african_ultrachat_amh	masakhane/african-ultrachat (Amharic train)	multi-turn chat
amharic_dolly	israel/AmharicDolly	machine-translated Dolly (instruction + optional context -> response)
amharic_alpaca	israel/AmharicAlpaca	machine-translated Alpaca; rows flagged error_suspicion removed
amharic_storyqa	israel/AmharicStoryQA_instruction-tuning (amharic)	story reading-comprehension, multiple choice and generation
muller_conversations	Muller01/amharic-conversations	multi-turn chat (system turns dropped)
daily_dialog_amh	abdukuzi45/amharic-daily-dialog-cleaned	long multi-turn everyday dialogue
kakugo_amh	ptrdvn/kakugo-amh	synthetic chat; English <think> blocks removed; rows that are not Amharic were filtered out
code170k_amh	michsethowusu/Code-170k-amharic	machine-translated code Q&A chat
addisai_alpaca, addisai_finetome	addisai/FreedomIntelligence-alpaca-amharic, addisai/FineTome-single-turn-dedup-amharic	machine-translated chat (50k sample each)
amharic_maternal_health	israel/AmharicMaternalHealth	health conversations; English-only conversations filtered out
tadesse_amharic_qa	Tadesse/amharic_qa	context question answering
roleplay_amh	jojo-ai-mst/Roleplay-Amharic	translated GPTeacher role-play (Amharic columns)
aya_human_amh	Henok/aya_amharic_dataset	human-written Aya instructions
addisgpt_amh	AddisGPT/AddisGPT-Amharic-Instruction	verified Amharic instructions
amharic_srh_qa	Walelign/Amharic_SRH_Dataset	sexual & reproductive health Q&A (CC-BY-NC)
afriinstruct_*	KoseiUemura/AfriInstruct-Data-amh (train)	Aya-style mix; FLORES removed; each sub-source capped at 10k
habtew_mt	habtew/english-amharic-translation (40k sample)	en<->am sentence pairs, with Walia's own MT instructions
afridocmt_amh	israel/AfriDocMT-instructions	document-level health translation (Amharic as source or target)

Row counts per source key:

source	rows
code170k_amh	175,045
afridocmt_amh	90,360
addisai_alpaca	49,809
addisai_finetome	49,641
amharic_alpaca	46,547
habtew_mt	39,507
kakugo_amh	29,045
walia_amharic_mezmur_completion	17,888
amharic_dolly	14,991
walia_amharic_title_generation	14,771
walia_afrisent	11,496
walia_amharic_mt amh-eng	10,997
walia_amharic_mt eng-amh	10,997
walia_amharic_spellcheck	10,852
walia_xlsum_summerization	10,698
walia_xlsum_reverse_summerization	10,695
walia_masakhaner	10,244
afriinstruct_SODA_inst_T	10,000
afriinstruct_AfriSenti	9,999
afriinstruct_Wiki_split_inst_T	9,998
walia_masakhanews	7,628
daily_dialog_amh	6,834
afriinstruct_Xlel_wd_inst_T	6,586
afriinstruct_MasakhaNEWS	6,225
afriinstruct_XL_Sum	5,746
walia_amharic_mezmur_generation	5,039
muller_conversations	5,000
african_ultrachat_amh	4,529
afriinstruct_HotpotQA_T	3,785
walia_amharic_poem_completion	3,776
amharic_storyqa	3,426
multiturn_combined_amh	3,375
tadesse_amharic_qa	2,617
walia_amharicqa	2,317
walia_amharic_story_generation	1,974
roleplay_amh	1,923
afriinstruct_NQ_Open_T	1,830
afriinstruct_CNN_Daily_Mail_T	1,228
aya_human_amh	1,192
walia_amharic_poem_generation	1,126
afriinstruct_MAFAND	1,035
amharic_maternal_health	1,011
wikiconversations_amh	1,010
amharic_srh_qa	991
afriinstruct_Flan_CoT_submix_T	916
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_Mintaka_inst_T	200
afriinstruct_AfriSenti_inst	190
afriinstruct_PIQA_T	170
afriinstruct_Dolly_v2_T	140
afriinstruct_Adversarial_QA_T	120
walia_amharic_zefen_generation	116
afriinstruct_Flan_Coqa_T	70
afriinstruct_Flan_lambada_T	50
afriinstruct_WIKI_QA_T	34
afriinstruct_MasakhaNEWS_inst	30

Known limitations (please read)

• Walia's English-script instructions are mixed in only where Walia itself has them (e.g. amharic_mt uses "English" inside Amharic instructions). No new templates were written for this dataset; Walia's own instructions are used as-is.
• amharic_alpaca and amharic_dolly are machine-translated and may contain translation errors. amharic_mt amh-eng outputs are English.
• Chat conversations with a decoding-collapse (a word repeated over and over) were removed from the chat sources (331 UltraChat conversations). The same check was not applied to lyrics/poems because they repeat on purpose; a small number of lyrics rows are still repetitive.
• Chat sources were also filtered for non-Amharic rows: conversations with a literal "None" answer, or with almost no Ethiopic script (except code chat), were dropped. This removed about a fifth of kakugo_amh and about 40% of amharic_maternal_health (English answers). Counts are in stats.json in the build output; the filter is a script-share heuristic, not a language-ID model.
• Not used on purpose: israel/AmharicMT (web-crawl pairs with hallucinated translations), FLORES-derived data (benchmark) and the test/validation splits of source datasets. afriinstruct_* is a mix of the same upstream corpora as parts of Walia (AfriSenti, MasakhaNEWS, XL-Sum), so some content overlaps; exact duplicates were removed but paraphrased overlap may remain.
• afridocmt_amh and habtew_mt are sentence/document translation pairs of mixed quality (religious and web text); habtew_mt outputs for amh-eng are English.
• amharic_srh_qa is CC-BY-NC-4.0, so the dataset as a whole should not be used commercially without removing it.
• 7,379 exact duplicate rows were removed in the Walia/Alpaca/Dolly part; near-duplicates and overlap between sources may remain.
• The Walia test split was not used, so it can still serve as an evaluation set. The Walia validation split was used as training data; validation here is a random per-task 0.5% holdout of everything.
• Some source repos are private, so this repo should stay private unless their licenses and owners allow release. Licenses differ per source (Alpaca- and Dolly-derived data carry their upstream terms); check each before commercial use.