引導式測驗 · 中等 等級
評估驅動開發測驗
使用代表性案例、基準、失敗分析、回歸測試和版本化提示來指導 LLM 申請的變更。
問題 1 的 8
評估驅動開發借鑒了哪些已建立的軟體實踐的結構?
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。
- 評估驅動開發借鑒了哪些已建立的軟體實踐的結構?
- Test-driven development, where a test is written before the code that satisfies it
- Waterfall project management
- Agile daily standups
- Object-oriented design patterns
- 為什麼深入研究說如果沒有評估套件,即時更改可能會有風險?
- Because prompt changes always cost more in API fees
- Because rewording one instruction to fix one failure can silently break a different one, and this might not surface until production
- Because prompts cannot be version-controlled
- Because models ignore prompt changes entirely
- 在模型切換範例中,評估套件揭示了更新、更快的模型的哪些內容?
- It was cheaper but had worse latency
- It failed to run at all
- It scored lower on factual accuracy for financial documents despite being faster
- It required a different programming language
- 團隊應該如何處理其評估套件中已確認的現實失敗?
- Keep every reported output forever without review
- Add only failures that are easy to automate
- Delete the case after one passing run
- Add a reproducible case when it reflects intended behavior, and revise or retire it if requirements change
- 為什麼在比較基準和候選提示時,總通過率可能會產生誤導?
- It can hide a specific case where the candidate got worse even while overall scores improved
- Aggregate numbers are always calculated incorrectly by eval tools
- Aggregate scores cannot be computed for LLM outputs
- Because only judge models produce aggregate scores
- Does eval-driven development require a large test suite before a prototype can be evaluated?
- Yes; every possible input must be listed
- No; start with a small set of important representative cases and grow it as evidence accumulates
- Yes, but only for models with tools
- No; manual demos replace all repeatable checks
- According to the technical insight, what three things does a working eval-driven workflow typically separate?
- The model provider, the API key, and the billing account
- The frontend, backend, and database layers
- The test cases, the scoring method, and a comparison report between baseline and candidate
- The prompt language, model size, and temperature setting
- Why can version-controlling prompts and eval suites together help teams investigate a regression?
- It reduces cloud storage costs
- Because version control systems require it for prompts specifically
- Because it makes prompts run faster at inference time
- It can link a changed prompt to the corresponding eval results in version history