---
base_model: google/gemma-3-1b-it
language:
- en
- sw
library_name: transformers
license: gemma
tags:
- swahili
- translation
- conversational
- gemma
- gemma3
- fine-tuned
pipeline_tag: text-generation
---

# Swahili Gemma 1B

A fine-tuned Gemma 3 1B instruction model specialized for **English-to-Swahili translation and Swahili conversational AI**. The model accepts input in both English and Swahili but outputs responses exclusively in Swahili.

## 📊 Translation Performance

![Translation Performance Comparison](swahili_gemma_ascending_chart.png)

### Model Comparison

| Model | Parameters | BLEU | chrF++ | Efficiency* |
|-------|------------|------|--------|-------------|
| Gemma 3 4B | 4B | 10.9 | 44.1 | 2.7 |
| **Swahili Gemma 1B** | **1B** | **27.6** | **56.8** | **27.6** |
| Gemma 3 27B | 27B | 29.4 | 60.0 | 1.1 |
| GPT-5 Mini | ~8B | 31.8 | 62.4 | 4.0 |
| Gemini 2.0 Flash | Large | 35.6 | 64.6 | N/A |

*Efficiency = BLEU Score / Parameters (in billions)

### Key Performance Insights

🎯 **Efficiency Leader**: Achieves the highest BLEU-to-parameter ratio (27.6 BLEU per billion parameters)  
🚀 **Size Advantage**: Outperforms Gemma 3 4B (4x larger) by 153% on BLEU score  
💎 **Competitive Quality**: Achieves 94% of Gemma 3 27B performance with 27x fewer parameters  
⚡ **Practical Deployment**: Runs efficiently on consumer hardware while maintaining quality  

### Evaluation Details

- **Dataset**: FLORES-200 English→Swahili (1,012 translation pairs)
- **Metrics**: BLEU (bilingual evaluation understudy) and chrF++ (character F-score)
- **Evaluation**: Zero-shot translation performance

## 🚀 Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("CraneAILabs/swahili-gemma-1b")
tokenizer = AutoTokenizer.from_pretrained("CraneAILabs/swahili-gemma-1b")

# Translate to Swahili
prompt = "Translate to Swahili: Hello, how are you today?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100, temperature=0.3)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```

## 🌍 Language Capabilities

- **Input Languages**: English + Swahili
- **Output Language**: Swahili only
- **Primary Focus**: English-to-Swahili translation and Swahili conversation

## 🎯 Capabilities

- **Translation**: English-to-Swahili translation
- **Conversational AI**: Natural dialogue in Swahili
- **Summarization**: Text summarization in Swahili
- **Writing**: Creative and informational writing in Swahili
- **Question Answering**: General knowledge responses in Swahili

## 💻 Usage Examples

### Basic Translation

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CraneAILabs/swahili-gemma-1b")
tokenizer = AutoTokenizer.from_pretrained("CraneAILabs/swahili-gemma-1b")

# English to Swahili translation
prompt = "Translate to Swahili: Good morning, how did you sleep?"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        inputs.input_ids,
        max_length=128,
        temperature=0.3,
        top_p=0.95,
        top_k=64,
        repetition_penalty=1.1,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

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

### Swahili Conversation

```python
# Direct Swahili conversation
prompt = "Hujambo! Je, unaweza kunisaidia leo?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100, temperature=0.3)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```

### Using the Pipeline

```python
from transformers import pipeline

# Create a text generation pipeline
generator = pipeline(
    "text-generation",
    model="CraneAILabs/swahili-gemma-1b",
    tokenizer="CraneAILabs/swahili-gemma-1b",
    device=0 if torch.cuda.is_available() else -1
)

# Generate Swahili text
result = generator(
    "Translate to Swahili: Welcome to our school",
    max_length=100,
    temperature=0.3,
    do_sample=True
)
print(result[0]['generated_text'])
```

### Ollama Usage

```bash
# Run the recommended Q4_K_M quantization
ollama run crane-ai-labs/swahili-gemma-1b:q4-k-m

# Try different quantizations based on your needs
ollama run crane-ai-labs/swahili-gemma-1b:q8-0    # Higher quality
ollama run crane-ai-labs/swahili-gemma-1b:q4-k-s  # Smaller size
ollama run crane-ai-labs/swahili-gemma-1b:f16     # Original quality
```

### Available Quantizations

| Quantization | Size | Quality | Use Case |
|-------------|------|---------|----------|
| `f16` | ~1.9GB | Highest | Maximum quality inference |
| `f32` | ~3.8GB | Highest | Research & benchmarking |
| `q8-0` | ~1.0GB | Very High | Production with ample resources |
| `q5-k-m` | ~812MB | High | Balanced quality/size |
| `q4-k-m` | ~769MB | Good | **Recommended** for most users |
| `q4-k-s` | ~745MB | Good | Resource-constrained environments |
| `q3-k-m` | ~689MB | Fair | Mobile/edge deployment |
| `q2-k` | ~658MB | Lower | Minimal resource usage |

## 💡 Generation Parameters

Recommended settings for optimal results:

```python
generation_config = {
    "temperature": 0.3,      # Focused, coherent responses
    "top_p": 0.95,          # Nucleus sampling
    "top_k": 64,            # Top-k sampling
    "max_length": 128,      # Response length limit
    "repetition_penalty": 1.1,  # Reduces repetition
    "do_sample": True,
    "pad_token_id": tokenizer.eos_token_id
}
```

## 🔗 Related Models

- **GGUF Quantizations**: [CraneAILabs/swahili-gemma-1b-GGUF](https://huggingface.co/CraneAILabs/swahili-gemma-1b-GGUF) - Optimized for llama.cpp/Ollama
- **LiteRT Mobile**: [CraneAILabs/swahili-gemma-1b-litert](https://huggingface.co/CraneAILabs/swahili-gemma-1b-litert) - Mobile deployment
- **Ollama**: [crane-ai-labs/swahili-gemma-1b](https://ollama.com/crane-ai-labs/swahili-gemma-1b) - Ready-to-run with Ollama

## 🎨 Use Cases

- **Language Learning**: Practice English-Swahili translation
- **Cultural Preservation**: Create and document Swahili content
- **Educational Tools**: Swahili learning assistants
- **Content Localization**: Translate materials to Swahili
- **Conversational Practice**: Improve Swahili dialogue skills
- **Text Summarization**: Summarize content in Swahili

## ⚠️ Limitations

- **Language Output**: Responds only in Swahili
- **Factual Knowledge**: General knowledge only, not trained on specific factual datasets
- **No Coding/Math**: Not designed for programming or mathematical tasks
- **Context Length**: Limited to 4,096 tokens for optimal performance
- **Specialized Domains**: May require domain-specific fine-tuning

## 📄 License

This model is released under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). Please review the terms before use.

## 🙏 Acknowledgments

- **Google**: For the Gemma 3 base model, support and guidance.
- **Community**: For Swahili language resources and datasets
- **Gilbert Korir (Msingi AI, Nairobi, Kenya)**
- **Alfred Malengo Kondoro (Hanyang University, Seoul, South Korea)**

## Citation

If you use this model in your research or applications, please cite:

```bibtex
@misc{crane_ai_labs_2025,
    author    = {Bakunga Bronson and Kato Steven Mubiru and Lwanga Caleb and Gimei Alex and Kavuma Lameck and Roland Ganafa and Sibomana Glorry and Atuhaire Collins and JohnRoy Nangeso and Tukamushaba Catherine},
    title     = {Swahili Gemma: A Fine-tuned Gemma 3 1B Model for Swahili conversational AI},
    year      = {2025},
    url       = {https://huggingface.co/CraneAILabs/swahili-gemma-1b},
    organization = {Crane AI Labs}
}
```


---

**Built with ❤️ by Crane AI Labs**

*Swahili Gemma - Your helpful Swahili AI companion*