Axiveri/Africlaude-7B API & Inference Endpoint

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

Table
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 Comparison

Table
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

Full Score Heatmap

Domain Mastery Radar

Fine-Tuning Impact vs Base Model

Fine-Tuning Impact

Score Progression Over 10 Training Cycles

Training Progression

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

Table
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

Table
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

python
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)

python
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

markdown
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

markdown
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

bibtex
@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

@Africlaude · @Axiveri

AfriclaudeBench v1.0 | Axiveri | All models tested live from HuggingFace

Africlaude-7B

Get help setting up a custom Dedicated Endpoints.

README