BeardedMonster/africomet-translation

AfriCOMET translation prompts

Known issue in v4, fixed in the next build. output_structure lists only the language of input in <|input_lang|>. The correct field lists every language of the user turn in reading order, e.g. an Igbo instruction before Efik text is <|input_lang|><ibo><efi>... (<efi><ibo> when the instruction is appended), and <|input_lang|><efi> only when the instruction and the text are in the same language. instruction_language and instruction_position are in every row, so the right field can be computed from them (rl.structure.input_lang_tags); the next build writes it directly. Prompts (prompt, text) are not affected.

10,000,003 translation prompts for RLHF with AfriCOMET as the reward. Each row is a passage of 1-5 consecutive sentences in source_lang plus an instruction asking for a translation into targ_lang. 0 rows (0.0%) also carry response, an aligned human reference translation (so reference-based AfriCOMET can be used on them); for the rest response is empty and the reference-free (QE) variant scores the policy's output against input.

No language model wrote any text here. Passages come from the public corpora listed below; instructions come from hand-written template banks.

Fields

field	meaning
instruction	the request, written in instruction_language, naming the target language
input	the source passage
response	aligned reference translation in targ_lang (empty when the source has no aligned text in one of our languages)
source_lang, targ_lang	eng pcm yor hau ibo fon ewe twi efi ibb urh. When response is set, targ_lang is the language of that reference
prompt	instruction and input joined in the order given by instruction_position (prepend / append)
instruction_language, template_confidence	language of the instruction (eng pcm yor hau ibo twi ewe) and how reliable that wording is (high/medium/low)
n_sentences, src_tokens	sentences in input; length in Davlan/afro-xlmr-large-76L tokens (capped at 256; response at 480; AfriCOMET truncates each segment at 512)
source_dataset, source_license, source_kind	provenance. kind: curated news encyclopedia web bible jw mined speech_text
africomet_supported	both languages are in AfriCOMET-1.1's encoder language list (Ibibio and Urhobo are not)

How it was built

• Wide sweep of Hugging Face. Every dataset tagged with, or named after, Pidgin, Efik, Ibibio, Urhobo, Fon, Twi/Akan or Ewe was inspected without downloading it: the first rows were read through the dataset-server and every text column was language-identified (GlotLID), so the right column is found by content, not by the dataset card. Parallel datasets contribute our language's column as the source; the aligned column becomes response when it is English or another of our languages, and is dropped otherwise (e.g. Ewe-French gives Ewe text only).
• Excluded after checking provenance: LLM-generated sets (e.g. pristine-twi, ghok-parallel-en-twi, pidgin-corpus-synth), a Twi set whose Twi side is NLLB output, translated task data (QA, sentiment, Alpaca), speech/TTS snippets, Cameroonian Pidgin, FLORES-only collections (xP3x), and FineWeb-2's 23 GB 'Pidgin' (misidentified English).
• Mined pair sets are only read near the top. Their NLLB similarity score is sorted descending and quality collapses past rank ~200k (mojibake, random names), so only rows above a score floor, capped per set, are used, and both sides are language-identified.
• Two cleaning profiles. Hausa, Yoruba, Igbo and English: strict (numbers and scripture references stripped, rule filters, plus a per-language prose classifier for Yoruba, Igbo and English that removed most machine-translated web text; only 13% of raw Yoruba and 24% of raw Igbo survived). Pidgin, Efik, Ibibio, Urhobo, Fon, Twi and Ewe: lenient (Bible verse numbers and references are kept; only mojibake, near-empty text, runaway repetition and mostly-English text are dropped), because their clean data is scarce.
• English is the augmenter. English passages without a reference fill the dataset and are routed to the other languages so each target language gets a similar number of rows.

Rows by source language

language	rows
hau	2,801,692
eng	2,371,986
afr	400,000
zul	400,000
sna	400,000
nya	400,000
xho	400,000
kin	400,000
som	400,000
swh	400,000
sot	397,674
orm	324,712
ibo	250,767
tsn	192,594
wol	118,386
yor	117,126
twi	88,385
ewe	75,734
pcm	27,620
fon	17,761
efi	15,566

Rows with a reference (response) by source language

language	rows
pcm	0
yor	0
hau	0
ibo	0
fon	0
ewe	0
twi	0
efi	0
afr	0
zul	0
sna	0
nya	0
xho	0
sot	0
kin	0
tsn	0
orm	0
wol	0
som	0
swh	0
eng	0

Rows by target language

targ_lang	rows
eng	726,472
pcm	463,695
yor	463,691
fon	463,690
hau	463,689
ewe	463,687
ibo	463,685
twi	463,685
afr	463,682
efi	463,682
zul	463,680
nya	463,679
sna	463,679
xho	463,678
kin	463,675
sot	463,668
orm	463,668
tsn	463,664
som	463,652
wol	463,652
swh	463,650

Rows by source kind

kind	rows
web	8,475,411
encyclopedia	1,281,848
bible	153,097
news	68,004
curated	21,643

Sources

Licenses differ per row; the dataset as a whole is only as permissive as its most restrictive rows.

source	rows
BeardedMonster/pretrain-new	4,233,366
HPLT/HPLT2.0_cleaned:hau_Latn/train-00001-of-00003.parquet	704,062
HPLT/HPLT2.0_cleaned:hau_Latn/train-00002-of-00003.parquet	621,586
HPLT/HPLT2.0_cleaned:hau_Latn/train-00000-of-00003.parquet	594,617
wikimedia/wikipedia:20231101.ha	198,712
cis-lmu/GlotCC-V1:hau-Latn	175,365
wikimedia/wikipedia:20231101.ig	151,706
HuggingFaceFW/fineweb-edu:sample/10BT/000	146,018
HuggingFaceFW/fineweb-edu:sample/10BT/001	144,821
HuggingFaceFW/fineweb-edu:sample/10BT/003	143,960
HuggingFaceFW/fineweb-edu:sample/10BT/002	143,897
HuggingFaceFW/fineweb-edu:sample/10BT/004	143,675
HuggingFaceFW/fineweb-edu:sample/10BT/005	142,261
HuggingFaceFW/fineweb-edu:sample/10BT/007	141,429
HuggingFaceFW/fineweb-edu:sample/10BT/006	141,372
wikimedia/wikipedia:20231101.en/00000	121,900
wikimedia/wikipedia:20231101.en/00001	116,960
allenai/c4:multilingual/c4-ha.tfrecord-00001-of-00008.json.gz	116,643
allenai/c4:multilingual/c4-ha.tfrecord-00000-of-00008.json.gz	115,603
allenai/c4:multilingual/c4-ha.tfrecord-00002-of-00008.json.gz	113,665
wikimedia/wikipedia:20231101.en/00003	113,313
allenai/c4:multilingual/c4-ha.tfrecord-00003-of-00008.json.gz	112,272
wikimedia/wikipedia:20231101.en/00002	112,249
wikimedia/wikipedia:20231101.en/00004	107,953
wikimedia/wikipedia:20231101.en/00006	99,011
wikimedia/wikipedia:20231101.en/00005	96,493
wikimedia/wikipedia:20231101.en/00007	85,381
HPLT/HPLT2.0_cleaned:ewe_Latn/train-00000-of-00001.parquet	42,815
allenai/c4:en/c4-train.00002	36,125
allenai/c4:en/c4-train.00004	36,092
allenai/c4:en/c4-train.00006	36,068
allenai/c4:en/c4-train.00001	36,054
allenai/c4:en/c4-train.00008	35,753
allenai/c4:en/c4-train.00007	35,620
allenai/c4:en/c4-train.00003	35,607
allenai/c4:en/c4-train.00005	35,579
allenai/c4:en/c4-train.00009	35,412
HPLT/HPLT2.0_cleaned:ibo_Latn/train-00000-of-00001.parquet	34,348
HPLT/HPLT2.0_cleaned:twi_Latn/train-00000-of-00001.parquet	33,683
allenai/c4:en/c4-train.00000	31,494
wikimedia/wikipedia:20231101.yo	28,282
HPLT/HPLT2.0_cleaned:yor_Latn/train-00000-of-00001.parquet	25,328
wikimedia/wikipedia:20231101.tw	19,515
masakhane/masakhanews:hau	18,544
masakhane/masakhanews:eng	17,489
masakhane/masakhanews:yor	11,329
masakhane/masakhanews:ibo	10,652
AfriSpeech/africa-corpus:Yoruba_yor_v2754.csv	10,250
AfriSpeech/africa-corpus:Igbo_ibo_v77.csv	10,248
AfriSpeech/africa-corpus:Efik_efi_v2570.csv	10,152
AfriSpeech/africa-corpus:Yoruba_yor_v911.csv	10,054
masakhane/masakhanews:pcm	9,990
AfriSpeech/africa-corpus:Éwé_ewe_v2259.csv	9,945
AfriSpeech/africa-corpus:Twi_twi_v1461.csv	9,535
AfriSpeech/africa-corpus:Yoruba_yor_v207.csv	9,471
cyanic-selkie/wikianc:ha	9,131
AfriSpeech/africa-corpus:Twi_twi_v1631.csv	8,755
AfriSpeech/africa-corpus:Twi_twi_v3439.csv	8,728
AfriSpeech/africa-corpus:Fon_fon_v813.csv	8,646
AfriSpeech/africa-corpus:Igbo_ibo_v1624.csv	8,593
AfriSpeech/africa-corpus:Éwé_ewe_v1613.csv	8,593
AfriSpeech/africa-corpus:Hausa_hau_v1614.csv	8,164
AfriSpeech/africa-corpus:Hausa_hau_v71.csv	7,843
AfriSpeech/africa-corpus:Pidgin,_Nigerian_pcm_v2516.csv	7,659
AfriSpeech/africa-corpus:Éwé_ewe_v3306.csv	7,651
cis-lmu/GlotCC-V1:yor-Latn	7,063
wikimedia/wikipedia:20231101.pcm	6,827
cyanic-selkie/wikianc:ig	5,738
allenai/c4:multilingual/c4-ig.tfrecord-00002-of-00004.json.gz	5,542
allenai/c4:multilingual/c4-ig.tfrecord-00001-of-00004.json.gz	5,531
allenai/c4:multilingual/c4-ig.tfrecord-00000-of-00004.json.gz	5,030
allenai/c4:multilingual/c4-ig.tfrecord-00003-of-00004.json.gz	4,958
cis-lmu/GlotCC-V1:ibo-Latn	4,547
allenai/c4:multilingual/c4-yo.tfrecord-00000-of-00002.json.gz	4,338
allenai/c4:multilingual/c4-yo.tfrecord-00001-of-00002.json.gz	4,242
HPLT/HPLT2.0_cleaned:fon_Latn/train-00000-of-00001.parquet	3,816
google/smol:smoldoc__en_yo	2,683
cyanic-selkie/wikianc:yo	2,678
google/smol:smoldoc__en_ha	2,662
AfriSpeech/africa-corpus:Efik_efi_v4541.csv	2,442

Licenses

license	rows
mixed (see pretrain-new card)	4,233,366
CC0-1.0	2,252,984
ODC-By-1.0	1,989,061
CC-BY-SA-3.0 / GFDL	1,262,328
unspecified (Bible translation; Bible-society copyright not verified)	153,097
AFL-3.0	68,004
CC-BY-SA-4.0	25,811
CC-BY-4.0	11,763
unknown (OPUS; mixed)	3,579
unspecified	10

Limitations

• Licenses are mixed and several are restrictive. CC-BY-NC rows (ghananlpcommunity/*, goldfish-models/fish-food, howard-nlp/ibom-mt, ...), share-alike rows, Bible and jw.org rows whose copyright is not cleared, and web rows that are Common Crawl derivatives. Filter on source_license / source_kind for a stricter release.
• Mined pair rows (source_kind=mined) have noisy alignment even above the score floor, so their response is a weaker reference than the Bible/curated ones. Most pair text is itself Bible/JW-style.
• Ibibio and Urhobo are not in AfriCOMET-1.1's language list; rows involving them have africomet_supported=false.
• Instruction templates are not native-reviewed and Fon, Efik, Ibibio and Urhobo have none (never an instruction language). Confidence: eng high, pcm medium, yor medium, hau medium, ibo medium, twi low, ewe low.
• Windows are 1-5 sentences but parallel/mined rows are single units, so n_sentences skews low there.
• FLORES-200 / SIB-200 sentences are standard evaluation sets: do not evaluate on FLORES after training on this.
• Efik and Ibibio sources mix n̄ and ñ; text is NFC-normalised only. No toxicity filtering was applied.

Reproduction

data-gen/africomet_translation/ in the SabiYarn repository: cli extract, cli clean, cli assemble.