Confam AI Adapter v1

Confam AI Adapter v1 is the first version of the Confam AI fine-tuning adapter, built on top of NCAIR1/N-ATLaS, a multilingual African language model based on Llama 3 8B.

The adapter is developed for Confam AI, a Nigerian-focused AI platform designed to provide helpful, culturally relevant and multilingual AI experiences.

🌐 Confam AI: https://confamai.saynet.ng/

🤗 Model: https://huggingface.co/Saynetng/confam-ai-adapter

🤗 Base Model: https://huggingface.co/NCAIR1/N-ATLaS

Version: v1
Status: Initial release

Model Details

Property	Details
Model name	Confam AI Adapter v1
Version	v1
Organization	Saynetng
Model type	Fine-tuning adapter
Base model	NCAIR1/N-ATLaS
Base architecture	Llama 3 8B
Task	Causal Language Modeling / Conversational AI
Languages	English, Hausa, Igbo, Yoruba
Platform	Confam AI
Format	Safetensors
Framework	Hugging Face Transformers / PEFT
Release	Version 1 / Initial release

About Confam AI

Confam AI Adapter v1 is the first publicly released version of the Confam AI adapter.

It extends the capabilities of the N-ATLaS multilingual African language model and is intended to provide a foundation for Confam AI's conversational AI experiences.

Confam AI focuses on building useful AI experiences for Nigerian and African users, with support for English and major Nigerian languages.

🌐 Visit Confam AI: https://confamai.saynet.ng/

Version History

v1 — Initial Release

Confam AI Adapter v1 is the first version of the adapter.

This release establishes the initial model foundation for Confam AI and builds upon the multilingual capabilities of N-ATLaS.

Future versions may introduce improvements to:

Instruction following
Conversation quality
Nigerian cultural understanding
Hausa, Igbo and Yoruba capabilities
Factual accuracy
Safety
Overall model performance

Note: v1 should be considered an initial release. Performance may improve in future versions as additional training, evaluation and optimization are performed.

Base Model

This adapter is fine-tuned from:

NCAIR1/N-ATLaS

N-ATLaS is a multilingual African language model based on Llama 3 8B, with support for English, Hausa, Igbo and Yoruba.

The Confam AI Adapter v1 builds on these capabilities for the Confam AI use case.

Supported Languages

Language	Support
🇬🇧 English	Supported
🇳🇬 Hausa	Supported
🇳🇬 Igbo	Supported
🇳🇬 Yoruba	Supported

Performance may vary between languages and tasks.

Intended Use

Confam AI Adapter v1 is intended for:

Conversational AI
Nigerian-focused AI applications
African language applications
Multilingual assistants
Educational applications
Language technology research
Translation and language transformation
AI-powered applications
Digital inclusion
English, Hausa, Igbo and Yoruba interactions

Training & Quantization

Confam AI Adapter v1 was fine-tuned from NCAIR1/N-ATLaS using parameter-efficient LoRA / PEFT fine-tuning with 4-bit NF4 quantization.

Training Configuration

Parameter	Configuration
Base model	NCAIR1/N-ATLaS
Fine-tuning method	LoRA / PEFT
Quantization	4-bit NF4
Double quantization	Enabled
4-bit loading	`load_in_4bit=True`
Quantization type	`bnb_4bit_quant_type="nf4"`
Double quantization	`bnb_4bit_use_double_quant=True`
Compute dtype	`float16`
4-bit compute dtype	`bnb_4bit_compute_dtype=torch.float16`
Training precision	FP16 (`fp16=True`)
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0.05
Learning rate	2e-4
LR scheduler	Cosine
Warmup steps	10
Optimizer	`paged_adamw_8bit`
Configured epochs	1
Maximum training steps	200
Effective batch size	8
Maximum sequence length	2048 tokens
Training records	643
Estimated dataset passes	~2.5 passes
Training hardware	NVIDIA T4
BF16	Not used

Quantization

The base N-ATLaS model was loaded using 4-bit NormalFloat (NF4) quantization with double quantization enabled.

load_in_4bit=True
bnb_4bit_quant_type="nf4"
bnb_4bit_use_double_quant=True
bnb_4bit_compute_dtype=torch.float16

The training run used float16 computation because the NVIDIA T4 hardware used for training does not provide the required BF16 support for this configuration.

LoRA Configuration

The adapter uses Low-Rank Adaptation (LoRA) to fine-tune a relatively small number of trainable parameters while keeping the majority of the underlying N-ATLaS model frozen.

LORA_RANK = 16
LORA_ALPHA = 32
LORA_DROPOUT = 0.05

LoRA Parameter	Value
Rank (r)	16
Alpha	32
Dropout	0.05

Optimization

The model was trained with a learning rate of 2 × 10⁻⁴ using a cosine learning-rate scheduler and 10 warmup steps.

Learning rate:  2e-4
Scheduler:      cosine
Warmup steps:   10
Optimizer:      paged_adamw_8bit

The paged_adamw_8bit optimizer was used to reduce optimizer memory requirements during fine-tuning.

Training Duration

Although num_train_epochs=1 was configured, the training run was step-limited to 200 steps.

With an effective batch size of 8 and 643 training records, the run corresponds to approximately 2.5 passes over the training records.

Training Parameter	Value
Training records	643
Effective batch size	8
Maximum steps	200
Configured epochs	1
Approx. dataset passes	~2.5

Sequence Length

The maximum sequence length was set to 2048 tokens.

N_ATLAS_MAX_SEQ = 2048

The training data was checked to ensure that no record exceeded the configured maximum sequence length.

Hardware

Training was performed using an NVIDIA T4 GPU.

Because the T4 was used for the training run, FP16 was used for computation rather than BF16.

Compute dtype:       float16
Training precision:  FP16
GPU:                 NVIDIA T4

The combination of 4-bit NF4 quantization, LoRA, and paged_adamw_8bit was used to make parameter-efficient fine-tuning feasible within the available GPU memory.

Training Method Summary

N-ATLaS (Llama 3 8B)
        │
        ▼
4-bit NF4 Quantization
        │
        ├── Double Quantization
        └── FP16 Compute
        │
        ▼
LoRA Fine-Tuning
        │
        ├── Rank: 16
        ├── Alpha: 32
        └── Dropout: 0.05
        │
        ▼
Cosine LR Scheduler
        │
        ├── Learning Rate: 2e-4
        └── Warmup: 10 steps
        │
        ▼
paged_adamw_8bit
        │
        ▼
200 Training Steps
        │
        ▼
Confam AI Adapter v1

Using the Adapter

Because this repository contains an adapter, it should be loaded together with the compatible N-ATLaS base model.

Installation

pip install -U transformers peft accelerate torch

Loading the Adapter

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

BASE_MODEL = "NCAIR1/N-ATLaS"
ADAPTER_MODEL = "Saynetng/confam-ai-adapter"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

# Load base model
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load Confam AI adapter
model = PeftModel.from_pretrained(
    model,
    ADAPTER_MODEL
)

messages = [
    {
        "role": "system",
        "content": "You are Confam AI, a helpful Nigerian AI assistant."
    },
    {
        "role": "user",
        "content": "Hello, how are you?"
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(
    prompt,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.7,
        top_p=0.9,
        do_sample=True
    )

response = tokenizer.decode(
    outputs[0],
    skip_special_tokens=True
)

print(response)

Model Architecture

Llama 3 8B
     │
     ▼
N-ATLaS
NCAIR1/N-ATLaS
     │
     │ Fine-tuning / LoRA Adapter
     ▼
Confam AI Adapter v1
Saynetng/confam-ai-adapter
     │
     ▼
Confam AI
confamai.saynet.ng

The Confam AI Adapter is not a standalone base model. It is designed to work with the compatible N-ATLaS base model.

Evaluation

Independent benchmark results for Confam AI Adapter v1 are not currently published.

Users are encouraged to evaluate v1 across:

English
Hausa
Igbo
Yoruba
Nigerian cultural context
Instruction following
Factual accuracy
Helpfulness
Safety
Hallucination
Conversation quality

Future versions may include expanded benchmarks and evaluation results.

Limitations

As v1, this model represents the initial version of the Confam AI adapter and should not be considered a final or fully optimized model.

Potential limitations include:

Hallucinations
Incorrect factual information
Uneven performance across languages
Bias
Cultural misunderstandings
Limited domain-specific knowledge
Sensitivity to prompting
Inconsistent responses
Performance limitations outside the training distribution

Users should independently verify important information generated by the model.

Responsible Use

Confam AI Adapter v1 should be used responsibly.

Developers should:

Verify important information.
Avoid using model outputs as a substitute for professional advice.
Test the model extensively before production deployment.
Evaluate performance across relevant Nigerian languages.
Monitor for harmful, biased or culturally inappropriate outputs.
Maintain human oversight for high-impact applications.

License and Attribution

Confam AI Adapter v1 is a derivative work based on NCAIR1/N-ATLaS.

Users should review and comply with the licensing and attribution requirements of the N-ATLaS model before using, modifying or redistributing this adapter.

For the complete terms, please refer to:

N-ATLaS: https://huggingface.co/NCAIR1/N-ATLaS

Citation

If you use Confam AI Adapter v1 in your project or research, please cite:

@misc{confamai_adapter_v1,
  title={Confam AI Adapter v1},
  author={Saynetng},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Saynetng/confam-ai-adapter},
  note={Version 1 fine-tuning adapter based on N-ATLaS}
}

For the underlying N-ATLaS model, please also refer to its official model card and citation.

Acknowledgements

Confam AI acknowledges the developers and contributors behind N-ATLaS and the broader Nigerian and African language AI ecosystem.

The Confam AI Adapter builds upon the work of the N-ATLaS project and its contributors.

Model Card Authors

Saynetng / Confam AI

🌐 https://confamai.saynet.ng/

🤗 https://huggingface.co/Saynetng

Confam AI Adapter — Version 1

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Saynetng/confam-ai-adapter

Base model

NCAIR1/N-ATLaS

Finetuned

(4)

this model

Saynetng
/

confam-ai-adapter