Add Synthesis Data Jcommonsenseqa using NeMo Data Designer with Nemotron-Personas-Japan seed

[SnowMasaya]

Overview

This notebook that implements a persona-seeded synthetic data generation pipeline for Japanese CommonsenseQA (jcommonsenseqa) using NeMo Data Designer and Nemotron Personas Japan.

---

What this PR adds

• A complete notebook for synthetic commonsense QA data generation
• Integration of Nemotron Personas Japan as structured generation seeds
• Generation of jcommonsenseqa-style multiple-choice questions with strict format constraints

---

Key design decisions

• Clear separation of roles
  • A large model (GPT-OSS-120B) is used exclusively for:
    • Synthetic data generation
    • Quality evaluation (judge)
  • Target models (e.g., Nemotron-Nano-9B-v2) are evaluated separately using llm-jp-eval
• Persona-aware but non-leaking seeds
  • Persona attributes are used to stabilize generation distribution
  • Seeds are not exposed directly in the generated questions

---