NVIDIA NCP-GENL 問題集 練習問題・模擬試験｜日本語版＆英語版対応

サンプル問題 1： A government agency is deploying an LLM for citizen services (benefits eligibility, tax questions, immigration status). Requirements: • Must serve all citizens equitably • Audit trail for all decisions • Ability to correct errors rapidly • Compliance with accessibility standards The model performs well in testing, but stakeholders worry about real-world fairness. Which deployment strategy best ensures responsible Al practices?

• A: Phased rollout starting with low-risk queries, expanding based on fairness metrics from each phase
• B: Parallel deployment with human agents handling sensitive cases while the LLM handles routine queries despite model biases
• C: Full deployment with a prominent feedback mechanism and weekly bias analysis of user interactions
• D: Blue-green deployment with ability to instantly rollback to previous versions if bias is detected

サンプル問題 2： Which of the following actions best represents a standard method for quantitatively evaluating the generative capability of a large language model (LLM)?

• A: Increasing the model's training data without measuring outcomes
• B: Relying exclusively on user feedback for all assessments
• C: Measuring model performance using metrics such as BLEU, ROUGE, and perplexity
• D: Modifying prompts to test new task capabilities

サンプル問題 3： You’re implementing a RAG system for a technical support chatbot with access to 10TB of documentation. Current challenges: • Documentation updates daily with version-specific information • Users often ask about error messages with slight variations • Need to handle multi-hop reasoning (e.g., ’error X usually means Y, and Y is fixed by Z') • Latency budget: 500ms end-to-end - Accuracy requirement: 95% for known issues Which RAG implementation best balances these requirements?

• A: Implement hierarchical indexing with sparse (BM25) for initial retrieval and dense embeddings for reranking, use incremental indexing for daily updates, add query expansion with LLM-generated variations, and implement iterative retrieval for multi-hop reasoning
• B: Build knowledge graph from documentation, use graph neural networks for retrieval, implement fuzzy matching for error variations, maintain separate indices per version, and use beam search for multi-hop paths
• C: Deploy hybrid sparse-dense retrieval in single stage, use vector database with HNSW index, implement document version tagging, generate multiple query embeddings, and limit to top-3 documents for latency
• D: Use dense-only retrieval with sentence transformers, implement semantic caching for common queries, rebuild entire index nightly, and use chain-of-thought prompting to handle multi-hop in single retrieval