「データ不足」の壁を越える：合成ペルソナが日本のAI開発を加速

AI development in Japan is hindered by a chronic lack of high-quality, culturally relevant training data. A study by NTT DATA demonstrates how synthetic data can overcome this "data wall" by generating large, privacy-preserving datasets from a small seed of proprietary data. Using NVIDIA's Nemotron-Personas-Japan dataset, the experiment significantly improved a model's accuracy from 15.3% to 79.3% on a legal document task, while also eliminating hallucinations. This approach allows for the creation of sovereign, domain-specific AI systems without compromising data privacy, enabling a more efficient development cycle by focusing on supervised fine-tuning.