Before You Start

Fifteen minutes here saves two to four weeks of duplicated groundwork.

Every serious African-language NLP project starts by re-doing work someone else already did. A new team collects sentiment data for Hausa that overlaps 60% with AfriSenti. A university group builds NER for Yoruba from scratch, months in, discovers MasakhaNER 2 covers it. A funder writes a proposal for parallel Amharic-English translation without checking LAFAND-MT. This chapter exists to stop that.

For every core task, this chapter answers the same four questions:

1. What already exists for African languages — datasets, models, benchmarks, per language, with an editorial opinion attached to each entry.
2. Fork or start fresh? — a decision tree that gets you to a defensible answer in five minutes.
3. What is this actually going to cost you — realistic effort estimates from the projects that did it before, in person-months and annotator-hours.
4. Where the canonical fine-tuning tutorial lives — one authoritative link out. We do not rewrite Hugging Face.

Available now

• Named Entity Recognition — the flagship reference chapter, built on MasakhaNER 1 and 2.
• Machine Translation — built on LAFAND-MT, MENYO-20k, FLORES-200, NLLB-200.
• Automatic Speech Recognition — built on Common Voice, FLEURS, NCHLT, MakerereNLP, Meta MMS, and Whisper.
• Sentiment analysis — built on AfriSenti (14 languages) and NaijaSenti.
• Hate speech and content safety — for teams working on a task whose failure modes are qualitatively different from classification-as-usual.
• Text-to-Speech — built on Meta MMS-TTS, VITS/XTTS/Kokoro architectures, and the voice-consent framing that TTS uniquely requires.
• Question Answering — built on AfriQA, TyDi QA, mDPR / mContriever retrieval, with cross-lingual open-domain framing.
• OCR and document AI — built on Tesseract, TrOCR, Kraken/Calamari, PaddleOCR, and community corpora across Latin, Ge'ez, Ajami, and other African scripts.

What's next

All seven initially-planned Before-You-Start pages are now available. Future additions — for languages, scripts, or task variants not yet covered — follow the same four-part structure: what exists / fork-or-fresh / cost / canonical link, with editorial opinions attached to each recommendation. See the contribution guide to propose one.

The one rule

If any entry on any of these pages is more than six months out of date, it is broken and needs fixing. Datasets get superseded. Models get deprecated. Recommendations that were right a year ago rot. Each page ships with a "last reviewed" date at the top, and the community owns keeping that date fresh.

If you find a page here past its six-month mark and need to know what has shipped since, the Finding current resources chapter names the primary Hugging Face organisations, archives, and workshops where the current state of the ecosystem lives.

What this is not

This chapter is not a fine-tuning tutorial. It is the map you consult before you decide whether to fine-tune, what to fine-tune, and against what. The tutorial itself lives one link out.

For the strategic reasoning behind this chapter — why the playbook adds a "Before You Start" section instead of writing more model-training content — see What this playbook is (and isn't).