FLAGSHIP measured on 9jaBench

NGPT-1 model card

A 7B instruction model across five Nigerian languages — Pidgin, Hausa, Yorùbá, Igbo and Nigerian English. Adapted from a strong open base on validated Nigerian data, aligned on real human preferences, and scored in the open. Every number here is real and measured; nothing is invented.

43.3

Pidgin chrF · frontier parity

38.0

clean chrF avg

+9.8

chrF over its base

Nigerian languages

Try NGPT-1 live Benchmarked on 9jaBench ↗

At a glance

Model: NGPT-1 — flagship of the NGPT line
Type: 7B-parameter instruction-tuned chat model
Base model: Qwen2.5-7B-Instruct — we adapt a strong open multilingual base; we do not pre-train a frontier model from scratch.
Languages: Nigerian Pidgin (flagship lane), Hausa, Yorùbá, Igbo, Nigerian English
API format: OpenAI chat-completions compatible · model id ngpt-1
Maker: Ranked Technologies Ltd · RC 9522220 · Lagos, Nigeria
License: coming soon — an open-weight NGPT-Community release with a published license is on the roadmap. No placeholder download links until weights are final.

Intended use

Chat and instruction-following in Nigerian Pidgin, Hausa, Yorùbá, Igbo and Nigerian English, including code-switching and local register.
Translation between English and the four Nigerian languages above.
Customer-support, document and citizen-facing language tasks for in-country deployment — see the domain builds (Banking, Energy, Government) delivered through LocalAI.

Out of scope / known limits

Grade-school math is a known weak spot — NGPT-1 is behind its base on AfriMGSM and we publish that loss.
Knowledge is roughly at base level — the fine-tune added language coverage, not new facts.
Not a safety-aligned production model on its own; the hardened, human-preference-aligned track is NGPT-Pro (in progress, via LocalAI).
Igbo is the hardest of the four languages and an area of active data work.

Benchmarks

NGPT-1 against its own base model, Qwen2.5-7B-Instruct, on the same public tasks. Wins and losses are both shown. Translation is the contamination-free (clean) chrF average; the raw average (39.4) is inflated by LaFAND test overlap, so 38.0 is the fair number. Full methodology and the contamination report live on the board.

TASK	BASE	NGPT-1	Δ
Translation chrF · 4 NG languages (clean)	28.2	38.0	+9.8
Knowledge AfriMMLU · accuracy %	45.0	44.2	-0.8
Reasoning AfriMGSM · accuracy %	8.0	5.0	-3.0

chrF is a character-level translation score (higher = closer to reference). AfriMMLU / AfriMGSM are accuracy on African-language knowledge and math sets. NGPT-1 is an unranked challenger on 9jaBench, openly behind the frontier overall — its honest flagship result is Pidgin parity (chrF 43.3, clean).

Full leaderboard & methodology on 9jaBench ↗

Training data & provenance

Base: the published Qwen2.5-7B-Instruct weights.
Instruction tuning: a validated ~121k-example Nigerian mix — translation pairs, instructions, and a deliberate anti-forgetting slice of math and general reasoning so the model retains base capabilities.
Preference alignment (NGPT-Pro track): real human preferences collected through 9jatesters — BVN-verified Nigerian raters, consented data.
Contamination note: part of the public Yorùbá translation test overlapped LaFAND training data, so the clean chrF figure (38.0) is reported as the fair number; the full contamination report is published on 9jaBench.
On continued pre-training: hundreds of millions of tokens of Nigerian web text cut held-out perplexity sharply but did not improve task scores, so it is not in NGPT-1.

How to run / self-host

NGPT-1 speaks the OpenAI chat-completions format, so any OpenAI-compatible client works against the live beta endpoint — point the base URL at NGPT and use model ngpt-1. The beta endpoint runs on Apple Silicon in Lagos, is rate-limited and may be briefly offline; no key is required yet.

curl

curl https://ngpt.ng/api/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ngpt-1",
    "messages": [
      {"role":"user","content":"Abeg explain blockchain for pidgin."}
    ]
  }'

python · openai sdk

from openai import OpenAI

client = OpenAI(
    base_url="https://ngpt.ng/api",
    api_key="not-needed",
)
r = client.chat.completions.create(
    model="ngpt-1",
    messages=[{"role":"user",
        "content":"Kọwaa ofe egusi n'Igbo."}],
)
print(r.choices[0].message.content)

Downloadable weights for offline self-hosting coming soon

An open-weight NGPT-Community release — weights plus a model card you can run or fine-tune locally — is on the roadmap and will publish here when finalised. There are deliberately no placeholder download links until the release is real. For on-premise or in-country enterprise deployment under data-residency rules, that track is NGPT-Pro, delivered through LocalAI.