CraneAILabs/ganda-gemma-1b

Ganda Gemma 1B: Specialized Luganda AI

Ganda Gemma 1B is a 1 billion parameter instruction-tuned model, built upon Google's Gemma 3 architecture and developed by Crane AI Labs. Its primary specialization is English-to-Luganda translation and Luganda conversational AI. A key differentiator is its remarkable efficiency, achieving a BLEU score of 6.99 per billion parameters, which is the highest among compared models.

Key Capabilities & Performance

• Exceptional Translation: Achieves a BLEU score of 6.99 and chrF++ of 40.32 on the FLORES-200 English→Luganda dataset, outperforming Gemma 3 4B (a model four times its size) by 535% in BLEU score.
• Competitive Quality: Demonstrates performance comparable to GPT-5 Mini, despite its significantly smaller size.
• Efficient Deployment: Designed for practical deployment, running efficiently on consumer hardware while maintaining high quality.
• Multifaceted Luganda AI: Beyond translation, it supports natural dialogue, summarization, creative writing, and question answering, all exclusively in Luganda.

Ideal Use Cases

• Translation Apps: Perfect for offline English-Luganda translation solutions.
• Language Learning: Aids in practicing Luganda with interactive feedback.
• Content Creation: Generates culturally relevant Luganda content for various media.
• Educational Tools: Functions as a Luganda learning assistant.

Limitations

• Output Language: The model generates responses only in Luganda.
• Context Length: Optimized for shorter conversational inputs (2048 tokens).
• Cultural Nuances: May not capture all intricacies of Luganda culture or regional dialect variations.