gemma2-2b-swahili-preview

Gemma2-2B-Swahili-Preview

Gemma2-2B-Swahili-Preview is a Swahili variation of the base language model Gemma2 2B fine-tuned on the Inkuba-Mono Swahili dataset, designed to enhance Swahili language understanding through monolingual training.

Model Details

• Developer: Alfaxad Eyembe
• Base Model: google/gemma-2-2b
• Model Type: Decoder-only transformer
• Language: Swahili
• License: Apache 2.0
• Fine-tuning Approach: Low-Rank Adaptation (LoRA)

Training Data

The model was fine-tuned on a focused subset of the Inkuba-Mono dataset:

• 1,000,000 randomly selected examples
• Total tokens: 60,831,073
• Average text length: 101.33 characters
• Diverse Swahili text sources including news, social media, and various domains

Training Details

• Fine-tuning Method: LoRA
• Training Steps: 2,500
• Batch Size: 2
• Gradient Accumulation Steps: 32
• Learning Rate: 2e-4
• Training Time: ~7.5 hours

Model Capabilities

This model is designed for:

• Swahili text continuation
• Natural language understanding
• Contextual text generation
• Base language modeling for Swahili

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("alfaxadeyembe/gemma2-2b-swahili-preview")
model = AutoModelForCausalLM.from_pretrained(
    "alfaxadeyembe/gemma2-2b-swahili-preview",
    device_map="auto",
    torch_dtype=torch.bfloat16
)

# Set to evaluation mode
model.eval()

# Example usage
prompt = "Katika soko la Kariakoo, teknolojia mpya imewezesha"
inputs = tokenizer(prompt, return_tensors="pt")
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)

Key Features

• Natural Swahili text continuation
• Strong cultural context understanding
• Efficient parameter updates through LoRA
• Diverse domain knowledge integration

Limitations

• Not instruction-tuned
• Base language modeling capabilities
• Performance varies across different text domains

Citation

@misc{gemma2-2b-swahili-preview,
  author = {Alfaxad Eyembe},
  title = {Gemma2-2B-Swahili-Preview: Swahili Variation of Gemma2 2B},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
}

Contact

For questions or feedback, please reach out through:

• HuggingFace: @alfaxadeyembe
• X : @alfxad