Gemma2-2B-Swahili-IT is a lightweight, efficient open variant of Google's Gemma2-2B-IT model, fine-tuned for natural Swahili language understanding and generation. This model provides a resource-efficient option for Swahili language tasks while maintaining strong performance.

Model Details

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

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: 31.58% accuracy
Fine-tuned Model: 38.60% accuracy
Improvement: +7.02%

Sentiment Analysis

Base Model: 84.85% accuracy
Fine-tuned Model: 86.00% accuracy
Improvement: +1.15%
Response Validity: 100%

Intended Use

This model is designed for:

Basic Swahili text generation
Question answering
Sentiment analysis
Simple creative writing
General instruction following in Swahili
Resource-constrained environments

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("alfaxadeyembe/gemma2-2b-swahili-it")
model = AutoModelForCausalLM.from_pretrained(
    "alfaxadeyembe/gemma2-2b-swahili-it",
    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: LoRA
Training Steps: 400
Batch Size: 2
Gradient Accumulation Steps: 32
Learning Rate: 2e-4
Training Time: ~8 hours on A100 GPU

Key Features

Lightweight and efficient (2B parameters)
Suitable for resource-constrained environments
Good performance on basic language tasks
Fast inference speed
Low memory footprint

Advantages

Resource Efficiency:
- Small model size (2B parameters)
- Lower memory requirements
- Faster inference time
- Suitable for deployment on less powerful hardware
Task Performance:
- Strong sentiment analysis capabilities
- Decent MMLU performance
- Good instruction following
- Natural Swahili generation

Limitations

Simpler responses compared to 9B/27B variants

Citation

bibtex

@misc{gemma2-2b-swahili-it,
  author = {Alfaxad Eyembe},
  title = {Gemma2-2B-Swahili-IT: A Lightweight Swahili Variant of Gemma2-2B-IT},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
}

Contact

For questions or feedback, please reach out through:

HuggingFace: @alfaxadeyembe
Twitter: @alfxad

Alfaxad/gemma2-2b-swahili-it

Gemma2-2B-Swahili-IT