引導式測驗 · 中等 等級

LLMOps 與 MLOps 測驗

將 LLM 特定的操作問題與提示、檢索、評估、成本和監控方面的共享 MLOps 實踐進行比較。

相關引導路徑Llmops 與 Mlops
問題 1 的 8

Which change can alter an LLM application's behavior without retraining foundation-model weights?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. Which change can alter an LLM application's behavior without retraining foundation-model weights?
    • A prompt template or retrieval index update
    • A dashboard color change only
    • A cloud region name in documentation
    • A model card font size
  2. Why version a retrieval corpus or index separately from the generator?
    • It makes the index part of model weights
    • It lets teams isolate whether changed evidence retrieval caused an output regression
    • It eliminates document access controls
    • It guarantees factual answers
  3. Which metric is specific to common LLM API cost tracking?
    • GPU occupancy only
    • Gini impurity
    • Input and output token usage
    • Number of class labels
  4. Why run evaluation cases after changing a prompt?
    • Evaluation is only for GPU kernels
    • Prompts cannot affect model responses
    • A prompt change retrains the model automatically
    • Prompt wording can alter outputs, tool use and refusal behavior
  5. What can model-based judging fail to detect?
    • Prompt version changes
    • Factual errors or evaluator bias shared with the candidate
    • Token counts
    • Retrieval index updates
  6. Which shared MLOps practice remains useful for LLM applications?
    • Using only training accuracy
    • Ignoring data governance
    • Staged release, monitoring and rollback
    • Always retraining full weights per request
  7. Why can a hosted model alias limit reproducibility?
    • Hosted models never return metadata
    • Aliases contain the full model weights
    • The prompt cannot be stored
    • Provider behavior may change behind an identifier
  8. Which security issue can arise from retrieved external documents?
    • Untrusted content can include prompt-injection instructions
    • Documents automatically retrain the model
    • Retrieval guarantees valid access permissions
    • Embeddings remove sensitive data