Machine Translation

Last reviewed: 2026-07-07.

Machine translation is the most operationally consequential African-language NLP task — it is what unlocks education, government, commerce, and health information in a speaker's own language. It is also the task where the largest published parallel corpora, the largest model releases, and the largest amount of well-intentioned but noisy web-scraped data collide. Getting this right for a new language starts with knowing what already exists.

What already exists

Four major bodies of work are the current reference points. Read this section top to bottom before deciding to build anything.

Community-curated parallel corpora

• LAFAND-MT / MAFAND-MT (Adelani et al., 2022) — 16 African-English news translation pairs, curated with a strong quality filter and human review. Repo and data on the Hugging Face Hub. This is the highest-quality Masakhane-lineage MT resource; treat it as the reference. Licence note: the corpus is released under CC-BY-4.0-NC (non-commercial); teams building commercial deployments must either restrict use, negotiate a licence with the corpus holders, or replace it with a commercially-permissive alternative.
• MAFAND-MT (Masakhane MT initial release) — the earlier community MT effort covering many African-English pairs. Superseded in coverage by LAFAND-MT for the languages both cover, but the repo remains the canonical directory of per-language Masakhane MT resources and links downstream.
• MENYO-20k (Adelani et al., 2021) — 20,100 Yoruba-English news pairs, curated with an explicit focus on tone marks and morphological consistency. The reference bilingual corpus for Yoruba MT.
• FLORES-200 (NLLB Team et al., 2022) — evaluation benchmark covering 200 languages including ~40 African. Use as an evaluation set, not training data — the splits are small on purpose. Available on the Hugging Face Hub.
• AfriDoc-MT — the Masakhane document-level MT corpus, funded by the Lacuna Fund. English paired with five African languages (Amharic, Hausa, Kiswahili, Yoruba, isiZulu) across two domains (Health and Technology), released with both document-level and sentence-level splits. The reference for teams whose deployment target is longer-form text — news articles, health information, government publications — rather than isolated sentences; LAFAND-MT's news-domain sentence splits do not cover this ground.
• AfriScience-MT — parallel scientific-domain MT corpus for English + six African languages (Amharic, Hausa, Luganda, Northern Sotho, Yorùbá, isiZulu) across eleven scientific domains including agriculture, biochemistry, and public health. Co-developed with expert science communicators and professional translators, and shipped with a domain-specific glossary alongside the parallel text. Apache-2.0 licensed. Also mirrored at dsfsi/afriscience_mt. The reference for teams building translation for scientific or technical content where terminology consistency matters as much as general fluency.
• AfriMTE-WMT2024 — the African MT quality-estimation challenge set from the WMT 2024 Metrics Shared Task. 13 African-language pairs (Darija–French, English↔Egyptian Arabic / Hausa / Igbo / Kikuyu / Luo / Somali / Swahili / Twi / isiXhosa / Yoruba, plus Yoruba↔English), 2,815 samples with source + machine translation + human reference + human quality scores. CC-BY-4.0. The reference for teams working on MT quality estimation or metric development for African languages; also useful as a small out-of-distribution eval set for teams shipping general MT.

Large web-scraped corpora — read with the retrievers' warnings

• NLLB Multilingual MT dataset (No Language Left Behind) — training data behind Meta's NLLB-200 model. Includes many African languages. Aggregated from many sources; some pairs are cleaner than others. Read the paper's data quality section before training on any given pair.
• JW300 (Agić & Vulić, 2019) — parallel religious text (Jehovah's Witness publications) across 300+ languages, including most African. Very high-coverage but severely domain-restricted; models trained only on JW300 fail on any secular content. Use for bootstrapping, never as the sole source.
• OPUS — the general OPUS aggregator hosts many African-language corpora with varied license and quality. Kreutzer et al. (2022) audited the low-resource end of these and found substantial mislabelling; the safe move is to spot-check any OPUS pair against native-speaker judgement before training.

Models — what has been trained on this data

• NLLB-200 and its larger variants (1.3B, 3.3B, 54B MoE) — Meta's dense and MoE MT models covering 200 languages. Strong 2022 baseline for out-of-the-box translation on any language in its coverage set; still widely used as a fine-tuning base, though newer options below are increasingly competitive. The distilled-600M variant is the practical fine-tuning starting point for teams without deep-pocket compute.
• SeamlessM4T v2 large — Meta's 2023 foundation model. 2.3B parameters, unified speech-to-speech + speech-to-text + text-to-speech + text-to-text translation + ASR across 101 speech-input languages and 96 text-input languages. Explicit African-language coverage includes Swahili, Yoruba, Igbo, Zulu, Shona, Somali, Nigerian Fulfulde. CC-BY-NC 4.0. Worth measuring against NLLB-200-distilled for your specific pair before committing to a fine-tune; SeamlessM4T often wins on speech-adjacent workflows and NLLB often wins on pure text.
• Cohere Aya-101 (Üstün et al. 2024) — the 13B foundation release covering 101 languages including many African languages; the honest "instruction-tuned multilingual LLM" pick for African-language MT via prompting. Note the successor caveat: the more recent Aya-Expanse (Dec 2024, CC-BY-NC-ND 4.0) supports only 23 languages, zero of them African (Cohere model card), so it is not a legitimate African-MT baseline despite its name; do not measure against it as one. Aya-101 is the Aya variant to keep in the candidate set.
• Toucan (Elmadany et al., 2024, ACL Findings) — UBC-NLP's explicitly Afrocentric MT model, released in 1.2B and 3.7B variants, fine-tuning Cheetah as base, supporting 156 African language pairs. This is the honest open Afrocentric MT reference in 2026 — larger African coverage than NLLB-200 among purpose-built African models — and should be measured against NLLB-200-distilled and SeamlessM4T v2 on your specific pair. Especially relevant for pairs NLLB-200 covers poorly or does not cover at all.
• TranslateGemma — Google's translation-focused Gemma 3 variants (4B / 12B / 27B params, 128K context), 55 languages including 15 African: Afar, Akan, Amharic, Bambara, Hausa, Igbo, Kinyarwanda, Lingala, Oromo, Shona, Somali, Swahili, Wolof, Yoruba, isiZulu. Available via Ollama; a strong African-language open translation model on Gemma-family bases as of 2026-Q3. Worth measuring against NLLB-200, SeamlessM4T v2, and Toucan on your specific pair.
• AfriqueLLM (Yu et al. 2026, ACL 2026 Main; GitHub) — McGill-NLP's continued-pretraining recipe applied to Llama 3.1, Gemma 3, and Qwen 3 bases for 20 African languages on 26B tokens. Nine released variants: AfriqueQwen (4B / 8B / 14B + Qwen 3.5 4B), AfriqueGemma (4B / 12B), AfriqueLlama-8B. CC BY 4.0. Framed by the paper as base models for downstream fine-tuning across tasks (MT, NER, QA, classification) — not translation-first, but a strong general-purpose African-language LLM base that competes with the Aya family on the languages both cover.
• AfriMT5 and community fine-tunes on the Hugging Face Hub under masakhane/ — smaller, per-language or per-family fine-tunes suited to constrained-compute deployment.
• mT5 and mBART-50 — general multilingual seq2seq baselines. Still useful as a fine-tuning starting point when NLLB is over-parameterised for the deployment.
• Meta MMS — while primarily a speech project, its supported-language list overlaps heavily with the African MT/ASR space and is worth knowing about for multimodal pipelines. SeamlessM4T v2 (above) is often the more integrated choice when speech-in-speech-out is the target.

Closed commercial APIs — worth knowing about as baselines

The playbook's editorial preference is for open models — self-hostable, licence-clear, community-inspectable. Closed commercial APIs are named here so teams can compare against them as baselines and make a deliberate build-vs-buy decision, not because they are the recommended default:

• Google Gemini — general-purpose multimodal LLM with translation as one capability. Claimed multilingual coverage is broad; per-language accuracy on African languages varies and must be measured.
• OpenAI GPT-4o and successors — general-purpose LLMs with translation capability. Same caveat.
• Anthropic Claude — general-purpose LLM with translation capability. Same caveat.
• Cohere Command — the closed commercial siblings of the Aya-family open releases; Cohere's African-language focus makes them worth measuring.

When closed APIs are the honest answer. For prototyping, for evaluating whether a task is even feasible on your target language, for one-off inference on a small test set, and for teams without infrastructure to self-host anything: closed APIs are often faster to try than deploying an open model. The tradeoffs — data sovereignty, licence terms permitting the provider to train on your inputs, reproducibility risk from model deprecation — are covered in the compute-poor chapter. Do not deploy production African-language systems on closed APIs without measuring per-language accuracy first; the "supported languages" list is a marketing claim, not an accuracy guarantee.

Editorial opinion. For a new project on a language covered by NLLB-200, NLLB-200-distilled remains a defensible fine-tuning starting point in mid-2026, but treat it as one of five candidates rather than the automatic default: measure NLLB-200 vs. SeamlessM4T v2 vs. Aya-101 vs. Toucan vs. TranslateGemma vs. AfriqueLLM (AfriqueGemma / AfriqueQwen / AfriqueLlama, whichever base matches your compute) on FLORES-200 dev + a native-speaker-verified small in-domain set before committing. Fine-tune the winner. Do not trust FLORES numbers without human review. If your language is not in NLLB-200, the task is genuinely research-grade — read the "cross-language transfer" chapter before scoping.

Honest expectation on scores. Public benchmarks show BLEU below 15 is common for African-language MT across NLLB, SeamlessM4T, and LLM-based approaches, and the gap between translating INTO English (easier) and out of a European lingua franca (harder) is substantial (Adelani et al., 2024; Aya-Expanse evaluation). Do not scope a project around a target of BLEU 30+ unless your specific pair has documented evidence to support that expectation.

Fork or start fresh?

Is your language pair covered by LAFAND-MT (16 African-English pairs)?
├── Yes → Use LAFAND-MT for the target-domain evaluation, fine-tune
│         NLLB-200-distilled starting from Meta's checkpoint.
│         Human-evaluate 200 random test outputs before publishing metrics.
└── No — is it in NLLB-200's 200-language coverage?
    ├── Yes → NLLB-200 out-of-the-box is your baseline. Collect a
    │         small (500-2000 sentence) native-speaker validation set
    │         in your target domain first — the NLLB numbers on
    │         out-of-distribution content can be misleading.
    └── No — genuinely uncovered.
        → Read the [long-tail language onboarding](../long-tail-language/index.md)
          chapter first for the realistic milestones and community /
          orthography / IP decisions that must come before data
          collection. Then use the [cross-language transfer](../cross-language-transfer/index.md)
          chapter for family-by-family pivot guidance, build a small
          parallel corpus (500-2000 sentences) as an evaluation set
          FIRST, and only then decide whether corpus creation, cross-
          language transfer, or reliance on an LLM API is the right
          investment.

What it will actually cost you

These are order-of-magnitude estimates. Your project will differ; the numbers below are the honest floor.

• Fine-tuning NLLB-200-distilled on an existing parallel corpus (LAFAND-MT or equivalent). Two to four person-weeks. One to five GPU-days for the distilled model, more for the larger checkpoints. Most of the elapsed time is evaluation and human review, not training.
• Curating a new 5-10k-sentence parallel corpus for training (news, government, health domains). Three to nine months elapsed; three to twelve person-months of translator work; ~$4-$15 per sentence at professional-translator rates, less when running through Masakhane-lineage community workflows with fair-rate volunteer coordination. Add one full person-month of lead effort for source selection, alignment, and quality review.
• Curating a new 500-2000-sentence evaluation-only set (the minimum viable for defensible human evaluation of a fine-tuned NLLB model). Four to eight weeks; one to two person-months of translator effort.
• Human evaluation of a new model on 200 randomly-sampled test outputs. Two person-weeks per evaluator; do this with at least two evaluators independently, then adjudicate disagreements. Do not trust a single-evaluator human eval.

Known limitations to watch for

• Domain drift is the dominant failure mode. A model trained on JW300 fails on news; a model trained on news fails on legal text; a model trained on formal text fails on social media. Report the training domain and the evaluation domain separately, and never publish a single BLEU or chrF number without stating both.
• BLEU is not the right primary metric for morphology-rich languages. Use chrF (character n-gram F-score) via sacrebleu. Report BLEU as a secondary number for comparison with prior work only. This is playbook-level editorial policy — see core principles.
• The corpus you can find is not the corpus you should train on. Kreutzer et al. (2022) audited widely-used web crawls of low-resource languages and found mislabelled, machine-translated, and off-language content at rates of 30-90% for many African pairs. Spot-check every corpus with a native speaker before training. This is not optional.
• Diacritic normalization changes results. Yoruba, Igbo, Vietnamese-style diacritic-carrying languages are especially sensitive. Publish your preprocessing pipeline explicitly. Never rely on a downstream user to guess it.
• Human evaluation is not optional for generative output. No automatic metric — including chrF, COMET, or BLEURT — replaces native-speaker judgement of adequacy and fluency. Budget human evaluation into every project from the start.
• Embedding-based metrics are a useful second opinion, not a headline. Alongside chrF as the primary and BLEU as the comparison-with-prior-work secondary, run one embedding-based metric on the same predictions to catch cases where character-level agreement misses semantic drift. For MT on African languages specifically, SSA-COMET (Wang et al., 2024) is the COMET-family variant fine-tuned on sub-Saharan African language pairs and is the honest first choice; the general COMET, BLEURT, and BERTScore are usable but under-calibrated on African languages. Newer general-purpose metrics — MetricX, xCOMET — are also worth spot-checking. Treat all embedding-based numbers as a comparison across model versions signal, not an absolute quality claim.
• Code-switching in training vs. test. Real African-language text mixes languages; most published parallel corpora are strict monolingual. If your target text is code-switched (Hausa-English, French-Wolof), the published corpus numbers overstate what you will get in deployment.
• Culturally-sensitive translation. Legal, medical, or religious content translated by a non-domain-expert translator is a liability, not an asset. Match translator expertise to content domain.

The canonical fine-tuning link

For fine-tuning NLLB-200 (or any seq2seq encoder-decoder), use the Hugging Face translation tutorial. It covers datasets loading, tokenisation for multilingual models, training loop, and evaluation with sacrebleu. The NLLB model card on the HF Hub has the specific fine-tuning guidance for that model, including how to specify the source and target language tokens correctly (this is a common footgun).

For sacrebleu itself (chrF and BLEU): github.com/mjpost/sacrebleu. Use the chrF metric and report the raw score, not a percentage.

Further reading

• LAFAND-MT paper (Adelani et al., 2022) — the reference methodology for community-curated African MT.
• NLLB paper (NLLB Team et al., 2022) — the definitive technical report on large-scale multilingual MT for low-resource languages.
• chrF paper (Popović, 2015) — for readers who want to understand why the playbook prefers chrF over BLEU.

Additional references

• Kreutzer et al., 2022 — quality of low-resource web crawls — required reading before touching OPUS or CC-100 low-resource pairs.
• MENYO-20k paper (Adelani et al., 2021) — the Yoruba reference corpus and the discussion of tone-mark handling.
• sacrebleu docs — the canonical evaluation implementation. Do not roll your own.