NaijaPidgin-Qwen3-4B

NaijaPidgin-Qwen3-4B 🇳🇬

A fine-tuned Qwen3-4B model that understands and speaks Nigerian Pidgin English (Naija) fluently.

Model Description

This model was fine-tuned from Qwen/Qwen3-4B using Unsloth with QLoRA on the naijaPidgin dataset.

Capabilities:

• Conversational Pidgin English
• Pidgin ↔ English translation
• Nigerian culture, proverbs, and everyday advice
• Code-switching between Pidgin and English
• Thinking mode (Qwen3 feature) for complex questions

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ahmadabdulnasir/NaijaPidgin-Qwen3-4B")
tokenizer = AutoTokenizer.from_pretrained("ahmadabdulnasir/NaijaPidgin-Qwen3-4B")

messages = [
    {"role": "system", "content": "You are a helpful assistant that speaks Nigerian Pidgin English. /no_think"},
    {"role": "user", "content": "How you dey?"},
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.7, top_p=0.8, top_k=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Thinking Mode

Remove /no_think from the system prompt to enable Qwen3's reasoning mode:

messages = [
    {"role": "system", "content": "You are a helpful assistant that speaks Nigerian Pidgin English."},
    {"role": "user", "content": "If I get 50,000 naira, wetin be di best investment?"},
]

With Ollama

Download the GGUF version and run locally:

ollama run ahmadabdulnasir/NaijaPidgin-Qwen3-4B

Training Details

• Base model: Qwen/Qwen3-4B
• Method: QLoRA (4-bit) with Unsloth
• LoRA rank: 32
• Learning rate: 2e-4
• Epochs: 3
• Hardware: Google Colab G4 GPU (96GB VRAM)
• Dataset: ahmadabdulnasir/naijaPidgin

Citation

@misc{naijapidgin_qwen3_2025,
  title={NaijaPidgin-Qwen3-4B},
  author={Ahmad Abdulnasir Shuaib},
  year={2025},
  url={https://huggingface.co/ahmadabdulnasir/NaijaPidgin-Qwen3-4B}
}