pip install olaverse

Flagship models

Three jobs, done properly: identify the language, restore its tone marks, generate from it. These are the models we maintain, version, and support for real use.

✍️ DiacNet — diacritic & tone restoration

Nigerian-language text without its tone marks is ambiguous to readers and nearly useless to machines. DiacNet puts them back.

• diacnet-1.0 — flagship multilingual ByT5 restorer (Yorùbá, Igbo, Hausa, and more)

from transformers import AutoTokenizer, T5ForConditionalGeneration

tok = AutoTokenizer.from_pretrained("olaverse/diacnet-1.0")
model = T5ForConditionalGeneration.from_pretrained("olaverse/diacnet-1.0")

inputs = tok("<yor> se eranko naa si gbo o?", return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))
# >> ṣé ẹranko náà sì gbọ́ ọ?

Prefix input with the target language tag — <yor>, <vie>, <ibo>, <hau>, <pol>, <tur>, <por>, <spa>, <fra>, <ita>.

🌍 LID — language identification

Most language detectors either don't cover Nigerian languages or confuse them with each other. Ours doesn't.

• lid-neural-25.2 — 25 languages, neural

from transformers import pipeline

lid = pipeline("text-classification", model="olaverse/lid-neural-25.2")
lid("Kedu ka i mere?")
# >> [{'label': 'ibo', 'score': 0.998}]

Use -25.2 for short text (queries, chat messages) and -25.1 for long-form passages.

⚡ MIST Generators — small models, one job each

• mist-qg-1.5b — multilingual question generation, 25+ languages (Qwen2.5-based)
• mist-tg-0.3b — 300M ByT5 title generator

Also on the Hub

Alongside the production tier we publish research and preview work — the MIST LLM line, the mist-encoder-base-ng foundation encoder and naija-embed-base embeddings, mist-reranker cross-encoders, OTK tokenizers, and the Prism vision models. These are open and usable, but they move faster and carry fewer guarantees. Browse all models →

🗺️ What's Next

• 📊 Published benchmarks for every production model — measured against real baselines, not claimed
• ✍️ DiacNet 1.1 — better Yorùbá & Igbo restoration, lower CER
• 🛠️ SDK CLI & REST inference (olaverse diacritize "se eranko naa")

Full roadmap in the docs.

🤝 Get Involved

We're a small lab with big ambitions. If you're working on African-language NLP, edtech, retrieval, or efficient specialist models — or want to use our models in production — reach out via olaverse.co.uk or open a discussion on any model.

Olaverse Lab — building AI that speaks your language.