Free NVIDIA Generative AI LLMs Associate Practice Questions

BERTScore computes similarity using contextual embeddings from a pre-trained model like BERT, allowing it to recognize semantic equivalence even when surface-level tokens differ. This is why semantically correct answers score well on BERTScore but poorly on BLEU, which requires exact n-gram overlap. Option A is not the primary mechanism. Option B describes a feature of some metrics but not what distinguishes BERTScore. Option D is incorrect as BERTScore considers all tokens through its embedding-based matching.

Strong diagonal attention patterns in later layers indicate the model is primarily attending to local context (nearby tokens) rather than forming long-range dependencies needed for multi-hop reasoning. For multi-hop tasks, later layers should show attention spanning across different sentences and their key entities. Option A would suggest a different problem (lack of specialization). Option C is actually normal behavior for learning sentence boundaries. Option D is expected in classification tasks and does not indicate a multi-hop reasoning failure.

Comprehensive red teaming involves domain experts systematically crafting adversarial prompts across multiple risk categories to proactively discover failure modes before deployment. This structured approach covers jailbreaks, social engineering, prompt injection, and other attack vectors. Standard benchmarks (B) measure general performance but are not designed to probe safety vulnerabilities. Beta deployment (C) is reactive rather than proactive and exposes real users to potential harms. Perplexity measurement (D) indicates model uncertainty but does not reveal specific safety-critical failure modes.