Africlaude-7B
Africa's own language model.
Built for African developers, Nigerian fintech engineers, and local innovation.
First model in the Africlaude series by Axiveri
Overview
Africlaude-7B is a 7-billion parameter language model built specifically for African developers, Nigerian fintech companies, and the African tech ecosystem. This is not a consumer chatbot. It is AI infrastructure — a foundation that African companies and developers build upon.
Most AI models are trained on Western-default data. They confuse kobo with cents. They have never seen a USSD flow, a Paystack webhook, or a CBN regulatory directive. Africlaude-7B was built to fix that.
What it understands natively:
- Nigerian English and Pidgin English
- USSD session management and AfricasTalking workflows
- Paystack, Flutterwave, and African payment API integrations
- CBN KYC tiers, licensing requirements, and compliance contexts
- Nigerian fintech fraud patterns and security awareness
- African developer coding patterns and tooling
Benchmark Results
Africlaude-7B was evaluated on AfriclaudeBench v1.0 — an open benchmark that loads and tests all models live from HuggingFace across 9 African domains. No estimates. No simulation. Real inference, real scores.
Overall Score Ranking
| Model | Overall Score |
|---|---|
| 🥇 Africlaude-7B (Axiveri) | 0.791 |
| LLaMA-3.1-8B (Meta) | 0.566 |
| Qwen2.5-Coder-7B (Base) | 0.555 |
| Mistral-7B-v0.3 (Mistral AI) | 0.533 |
Category Breakdown
| Category | Africlaude-7B | LLaMA | Mistral |
|---|---|---|---|
| Fintech | 0.86 | 0.36 | 0.43 |
| Security | 0.75 | 0.75 | 0.75 |
| Pidgin English | 1.00 | 1.00 | 0.00 |
| Identity | 1.00 | 0.00 | 0.00 |
| Fraud Detection | 1.00 | 1.00 | 1.00 |
| SQL/DB | 1.00 | 1.00 | 1.00 |
| USSD Engineering | 0.56 | 0.56 | 0.78 |
| Reasoning | 0.50 | 0.50 | 0.50 |
| CBN Compliance | 0.33 | 0.00 | 0.33 |
On Pidgin English — every competing model scored 0.00. Africlaude scored 1.00.
Full Heatmap
Domain Mastery Radar
Fine-Tuning Impact vs Base Model
Score Progression Over 10 Training Cycles
Peak score of 0.918 achieved at Cycle 9 NB3. Final released model scored 0.791 on AfriclaudeBench v1.0 against industry models.
Model Details
| Property | Value |
|---|---|
| Model Name | Africlaude-7B |
| Model ID | Axiveri/Africlaude-7B |
| Series | Africlaude (first model in the series) |
| Parameters | 7B |
| Format | Merged float16 — fully standalone, no external dependencies |
| License | Apache 2.0 |
| Languages | English, Nigerian Pidgin, Nigerian English |
| Developer | Emmanuel Ariyo |
| Organization | Axiveri |
Training Details
| Property | Value |
|---|---|
| Hardware | NVIDIA Tesla T4 x2 |
| VRAM | ~15.6 GB per GPU |
| Total GPU Hours | ~100 hours |
| Training Duration | 2 months |
| Training Cycles | 10 cycles |
| Fine-tuning Method | QLoRA (4-bit quantized LoRA) |
| LoRA Rank | r=16, alpha=32 |
| Framework | Unsloth + TRL + HuggingFace Transformers |
Training Data composition:
- Nigerian fintech workflows (Paystack, Flutterwave, CBN compliance)
- Pidgin English instruction pairs
- USSD session engineering
- African fraud detection and security awareness
- General reasoning (OpenR1-Math, GSM8K, MetaMath, NuminaMath)
- Code generation (CodeFeedback, OpenHermes)
- Financial instruction following (Finance-Alpaca)
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "Axiveri/Africlaude-7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float16,
device_map="auto"
)
messages = [
{"role": "user", "content": "How do I verify a Paystack webhook in Python?"}
]
text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
With Unsloth (Faster Inference)
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="Axiveri/Africlaude-7B",
max_seq_length=2048,
dtype=torch.float16,
load_in_4bit=True,
)
FastLanguageModel.for_inference(model)
Example Outputs
Pidgin English
User: Wetin be Paystack webhook?
Africlaude: Paystack webhook na like callback URL wey Paystack go use
send notification to your server when something happen — like when
customer pay, when payment fail, or when refund process. You go need
verify the signature for header to make sure na real Paystack send am,
not fake request.
Nigerian Fintech
User: Explain CBN KYC tiers for a Nigerian fintech
Africlaude: CBN KYC tiers in Nigeria:
Tier 1 (Basic): Maximum balance N300,000, daily transaction limit
N50,000. Requires BVN or NIN only.
Tier 2: Maximum balance N500,000, daily limit N200,000.
Requires BVN + one government-issued ID...
Known Limitations
- CBN Compliance (0.33) and Reasoning (0.50) are below target — addressed in the next training cycle
- AfriclaudeBench v1.0 has known scoring limitations; v2.0 is in development with harder, more African-specific questions
- Primarily optimised for Nigerian English and Pidgin; broader African language support (Yoruba, Hausa, Igbo) is on the roadmap
What's Next
Africlaude-7B is the first model in the Africlaude series — the learning and educational foundation for African developers and fintech teams. More powerful models with broader African language support, multimodal capabilities, and production-grade fintech engineering are in active development.
Follow @Africlaude and @Axiveri for updates.
Citation
@misc{africlaude7b2026,
title={Africlaude-7B: African-Standard Language Model for Fintech and Developer Contexts},
author={Emmanuel Ariyo},
organization={Axiveri},
year={2026},
url={https://huggingface.co/Axiveri/Africlaude-7B},
note={Fine-tuned language model. Evaluated on AfriclaudeBench v1.0
across 9 African domains. 10 training cycles on Tesla T4 x2.}
}
License
Apache 2.0 — free to use, modify, and build upon commercially. See LICENSE.
Built by Emmanuel Ariyo | Axiveri
AfriclaudeBench v1.0 | Axiveri | All models tested live from HuggingFace