mgbam/gaialab-naija-adapter-v0.1

GaiaLab Naija Adapter v0.1

GaiaLab Naija Adapter v0.1 is an experimental LoRA adapter trained on Qwen/Qwen2.5-0.5B-Instruct.

It explores Nigerian small-business communication, business writing, Nigerian English, and Nigerian Pidgin.

Intended use

The adapter is designed for research and experimentation involving:

• Nigerian small-business customer communication
• Professional business-writing assistance
• English-to-Nigerian-Pidgin translation
• Nigerian-Pidgin-to-English translation
• Explanation of common business terminology

Training data

• 100 curated examples
• 80 training records
• 20 validation records
• 5 task categories

The dataset includes:

• Customer-service responses
• Nigerian business terminology
• English-to-Nigerian-Pidgin translation
• Nigerian-Pidgin-to-English translation
• Business writing

Training configuration

• Base model: Qwen/Qwen2.5-0.5B-Instruct
• Fine-tuning method: LoRA
• Epochs: 3
• Training records: 80
• Validation records: 20

Training results

• Final training loss: 2.342
• Final evaluation loss: 2.106

These results confirm that the adapter-training pipeline completed successfully. They do not prove that the adapter is more accurate than the base model across all tasks.

Preliminary evaluation

Initial testing showed mixed results:

• Business-writing responses were generally clear and professional.
• Customer-service responses were polite but sometimes omitted requested details.
• Nigerian Pidgin translation quality was inconsistent.
• Some terminology responses contained factual errors.
• The model occasionally added information that was not present in the prompt.

This release should therefore be treated as a research prototype.

Limitations

The model may:

• Produce inaccurate information
• Change or omit details during translation
• Generate unnatural Nigerian Pidgin
• Misunderstand cultural or business context
• Add unsupported facts or explanations
• Hallucinate business, legal, tax, financial, or regulatory guidance

Human review is required before real-world use.

Loading the adapter

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

base_model_id = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_id = "mgbam/gaialab-naija-adapter-v0.1"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    dtype=torch.float16,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    adapter_id
)

model.eval()

Load the tokenizer from the base model because it contains the correct
Qwen chat template.

Model repository

Hugging Face:

mgbam/gaialab-naija-adapter-v0.1

Source code

GitHub:

oluwafemidiakhoa/gaialab-naija-assistant

Disclaimer

This is an experimental research release. Generated responses should be
reviewed before use in business, legal, financial, tax, medical, or
regulatory contexts.