Gemma2-27B-Swahili-IT

Gemma2-27B-Swahili-IT is a state-of-the-art open variant of Google's Gemma2-27B-IT model, fine-tuned for natural Swahili language understanding and generation. This model utilizes Quantized Low-Rank Adaptation (QLoRA) to achieve efficient fine-tuning while maintaining performance.

Model Details

• Developer: Alfaxad Eyembe
• Base Model: google/gemma-2-27b-it
• Model Type: Decoder-only transformer
• Language(s): Swahili
• License: Apache 2.0
• Finetuning Approach: QLoRA (4-bit quantization)

Training Data

The model was fine-tuned on a comprehensive dataset containing:

• 67,017 instruction-response pairs
• 16,273,709 total tokens
• Average 242.83 tokens per example
• High-quality, naturally-written Swahili content

Performance

Massive Multitask Language Understanding (MMLU) - Swahili

• Base Model: 22.81% accuracy
• Fine-tuned Model: 57.89% accuracy
• Improvement: +35.08%

Swahili Sentiment Analysis

• Base Model: 89.90% accuracy
• Fine-tuned Model: 90.00% accuracy
• Perfect response validity (100%)

Intended Use

This model is designed for:

• Natural Swahili text generation
• Question answering
• Content analysis
• Creative writing
• General instruction following in Swahili

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

# Configure 4-bit quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True
)

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("alfaxadeyembe/gemma2-27b-swahili-it")
model = AutoModelForCausalLM.from_pretrained(
    "alfaxadeyembe/gemma2-27b-swahili-it",
    quantization_config=bnb_config,
    device_map="auto",
    torch_dtype=torch.bfloat16
)

# Always set to eval mode for inference
model.eval()

# Example usage
prompt = "Eleza dhana ya uchumi wa kidijitali na umuhimu wake katika ulimwengu wa leo."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=500,
        do_sample=True,
        temperature=0.7,
        top_p=0.95
    )

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

Training Details

• Fine-tuning Method: QLoRA (4-bit quantization)
• Training Steps: 150
• Batch Size: 1
• Gradient Accumulation Steps: 64
• Learning Rate: 1.5e-4
• Training Time: ~10 hours on A100 GPU

Citation

@misc{gemma2-27b-swahili-it,
  author = {Alfaxad Eyembe},
  title = {Gemma2-27B-Swahili-IT: Swahili Variation of Gemma2-27b-it Model},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
}

Contact

For questions or feedback, please reach out through:

• HuggingFace: @alfaxad
• Twitter: @alfxad