Sunbird/Sunflower-14B
Sunbird AI's Sunflower-14B is a 14 billion parameter multilingual causal language model built on the Qwen3-14B architecture, specifically designed for Ugandan languages and English. This model excels at translation, text generation, and factual question-answering across 31 Ugandan languages, featuring a GRPO fine-tuned checkpoint for improved accuracy, fluency, and reduced hallucinations. It is optimized for educational, knowledge-access, and cross-lingual communication applications within Uganda.
Loading preview...
Sunflower-14B: Multilingual LLM for Ugandan Languages
Sunflower-14B, developed by Sunbird AI, is a 14 billion parameter multilingual causal language model based on the Qwen3-14B architecture. It is uniquely optimized for 31 Ugandan languages plus English, making it a crucial tool for regional language processing.
Key Capabilities & Differentiators
This model's latest release incorporates a GRPO (Group Relative Preference Optimization) fine-tuned checkpoint, significantly enhancing its performance compared to previous versions and other leading models:
- Superior Multilingual Translation: Achieves chrF scores of 0.419 (xx→eng) and 0.366 (eng→xx), outperforming Gemini 2.5 Pro and GPT-4o in these benchmarks for Ugandan languages.
- Enhanced Factual Q&A: Provides stronger factual correctness and improved answer grounding, especially in multilingual contexts.
- Improved Instruction Following & Consistency: Demonstrates more reliable instruction adherence and reduced repetitive or inconsistent outputs.
- Reduced Hallucinations: The GRPO fine-tuning specifically targeted and improved the model's ability to produce coherent and reliable responses.
Intended Uses
Sunflower-14B is ideal for applications requiring robust multilingual capabilities in Ugandan languages:
- Translation: Between English and Ugandan languages, and among different Ugandan languages.
- Text Generation: Creating content in various Ugandan languages.
- Multilingual Factual Question Answering: Accessing and disseminating knowledge in local contexts.
- Educational & Knowledge-Access: Supporting learning and information retrieval in diverse linguistic settings.
Training Insights
The model was trained on approximately 750 million characters of multilingual text, including digitized books, radio transcripts, web data (MADLAD-400, Common Crawl), and existing multilingual datasets. The training involved a multi-stage process: continued pretraining on Qwen3-14B, supervised fine-tuning, initial preference optimization, and the new GRPO fine-tuning for targeted improvements in factual correctness and multilingual reasoning.