CraneAILabs/ganda-gemma-1b-litert

Ganda Gemma 1B - LiteRT

LiteRT (formerly TensorFlow Lite) optimized version of Ganda Gemma 1B - a fine-tuned Gemma 3 1B instruction model specialized for English-to-Luganda translation and Luganda conversational AI.

This repository contains MediaPipe task bundles optimized for mobile deployment on Android and iOS devices.

📊 Translation Performance

FLORES-200 Evaluation Results

Our Ganda Gemma 1B model demonstrates strong performance in English-to-Luganda translation:

Metric	Score	Ranking
BLEU	6.99	2nd out of 5 models
chrF++	40.32	2nd out of 5 models

Model Comparison

Model	Parameters	BLEU	chrF++	Efficiency*
Gemma 3 4B	4B	1.1	20.05	0.28
Gemma 3 27B	27B	3.65	31.37	0.14
GPT-5 Mini	N/A	5.14	36.55	N/A
Ganda Gemma 1B	1B	6.99	40.32	6.99
Gemini 2.0 Flash	Large	7.94	43.38	N/A

*Efficiency = BLEU Score / Parameters (in billions)

Key Performance Insights

🎯 Efficiency Leader: Achieves the highest BLEU-to-parameter ratio (6.99 BLEU per billion parameters)
🚀 Size Advantage: Outperforms Gemma 3 4B (4x larger) by 535% on BLEU score
💎 Competitive Quality: Outperforms GPT-5 Mini by 36% on BLEU score with known parameter count
⚡ Practical Deployment: Runs efficiently on consumer hardware while maintaining quality

Evaluation Details

• Dataset: FLORES-200 English→Luganda (1,012 translation pairs)
• Metrics: BLEU (bilingual evaluation understudy) and chrF++ (character F-score)
• Evaluation: Zero-shot translation performance
• Model: ganda-gemma-1b checkpoint with GRPO enhancement

📱 Available Models

File	Size	Quantization	Use Case
ganda-gemma-1b-instruct.task	~978MB	FP16	Recommended - Instruction following format with MediaPipe bundling
ganda-gemma-1b.tflite	~973MB	FP16	Raw TFLite model - requires custom tokenizer integration

🚀 Quick Start

Android (MediaPipe)

import com.google.mediapipe.tasks.genai.llminference.LlmInference

// Load the model
val options = LlmInference.LlmInferenceOptions.builder()
    .setModelPath("/path/to/ganda-gemma-1b-instruct.task")
    .build()

val llmInference = LlmInference.createFromOptions(context, options)

// Generate response
val response = llmInference.generateResponse("Translate to Luganda: Good morning")
println(response)

iOS (MediaPipe)

import MediaPipeTasksGenAI

// Load the model
let options = LlmInference.Options()
options.modelPath = "/path/to/ganda-gemma-1b-instruct.task"

let llmInference = try LlmInference(options: options)

// Generate response
let response = try llmInference.generateResponse(inputText: "Translate to Luganda: Good morning")
print(response)

Web (MediaPipe)

import { LlmInference } from '@mediapipe/tasks-genai';

const llm = await LlmInference.createFromModelPath(
  '/path/to/ganda-gemma-1b-instruct.task'
);

const response = await llm.generateResponse('Translate to Luganda: Good morning');
console.log(response);

🌍 Language Capabilities

• Input Languages: English + Luganda
• Output Language: Luganda only
• Primary Focus: English-to-Luganda translation and Luganda conversation

📦 Model Variants Guide

1. ganda-gemma-1b-instruct.task (RECOMMENDED)

Best for: Most mobile applications - ready-to-use MediaPipe bundle

Input format: Natural instructions

Translate to Luganda: Hello, how are you?

MediaPipe formats as:

### Instruction:
Translate to Luganda: Hello, how are you?

### Response:

2. ganda-gemma-1b.tflite

Best for: Custom integrations requiring direct TFLite model access

Requirements: You need to handle tokenization manually Use case: Advanced users who want to integrate with custom tokenizers or frameworks

🎯 Capabilities

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

💡 Generation Parameters

Optimal settings for mobile deployment:

// JavaScript/Web
const response = await llm.generateResponse(prompt, {
  temperature: 0.3,
  topK: 40,
  randomSeed: 42
});

// Android
val response = llmInference.generateResponse(
    inputText = prompt,
    temperature = 0.3f,
    topK = 40,
    randomSeed = 42
)

// iOS
let options = LlmInference.Options()
options.temperature = 0.3
options.topK = 40
options.randomSeed = 42

let response = try llmInference.generateResponse(
    inputText: prompt,
    options: options
)

📱 Mobile Integration

Memory Requirements

• RAM: Minimum 3GB recommended for optimal performance
• Storage: ~1.2GB per task bundle
• CPU: ARMv8 or newer recommended

Performance Tips

1. Preload models during app initialization
2. Use appropriate quantization: FP16 provides good quality for mobile
3. Cache responses for repeated queries
4. Batch processing for multiple translations

🔗 Related Models

• Original Model: CraneAILabs/ganda-gemma-1b - Full precision HuggingFace model
• GGUF Quantizations: CraneAILabs/ganda-gemma-1b-GGUF - Optimized for llama.cpp/Ollama
• Ollama: crane-ai-labs/ganda-gemma-1b - Ready-to-run with Ollama

🎨 Use Cases

• Mobile Translation Apps: Offline English-Luganda translation
• Language Learning: Practice Luganda with instant feedback
• Cultural Apps: Create culturally aware Luganda content
• Educational Tools: Luganda learning assistants for mobile
• Offline AI: No internet required after model download
• Edge Computing: Run AI locally on mobile devices

⚠️ Limitations

• Language Output: Responds only in Luganda
• Mobile Resources: Requires significant RAM and storage
• Context Length: Optimized for shorter inputs on mobile
• Quantization: FP16 requires more memory than INT4/INT8
• Platform Support: Requires MediaPipe Tasks GenAI support

🛠️ Development Setup

Android

dependencies {
    implementation 'com.google.mediapipe:tasks-genai:latest.release'
}

iOS

// Add to Package.swift
.package(url: "https://github.com/google/mediapipe", from: "0.10.0")

Web

npm install @mediapipe/tasks-genai

📄 License

This model is released under the Gemma Terms of Use. Please review the terms before use.

🙏 Acknowledgments

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

Citation

If you use these LiteRT models in your research or mobile applications, please cite:

@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     = {Ganda Gemma: A Fine-tuned Gemma 3 1B Model for Luganda conversational AI},
    year      = {2025},
    url       = {https://huggingface.co/CraneAILabs/ganda-gemma-1b},
    organization = {Crane AI Labs}
}

---

Built with ❤️ by Crane AI Labs

Ganda Gemma - Your helpful Luganda AI companion, now on mobile!