引導式測驗 · 硬 等級

自我 RAG 和反思檢索測驗

回顧 Self-RAG 自適應檢索、反思標記、監督批評訓練、分段評分以及模型生成的支持判斷的限制。

相關引導路徑自 RAG 和反射檢索
問題 1 的 8

與針對每個輸入都進行檢索的基準模型相比,Self-RAG 學習到了什麼自適應決策?

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. 與針對每個輸入都進行檢索的基準模型相比,Self-RAG 學習到了什麼自適應決策?
    • Which operating system to install for every query
    • Whether retrieval is needed, including decisions as generation progresses
    • How to permanently remove all document sources
    • How to prove every generated statement without inspecting evidence
  2. Self-RAG 中的「反射令牌」是什麼?
    • 文件的加密金鑰
    • 使用者登入憑證
    • 檢索到的文件上的時間戳
    • 模型發出的特殊控制令牌,用於判斷相關性、支持性和有用性
  3. Self-RAG 的 IsSupported 反思標記預測了什麼?
    • Whether the model's hardware has enough disk space
    • Whether a support ticket has been paid
    • Whether a generated segment is supported by the retrieved evidence
    • Whether every source claim is independently proven true
  4. 通常如何訓練 Self-RAG 模型來產生反射標記?
    • 對標有反射標記的數據進行監督學習,通常從更強大的模型中提取
    • 純無監督聚類
    • 透過手動編碼規則而無需學習
    • 僅透過人類點擊進行強化學習
  5. 對每個段落發出“IsRelevant”批評有什麼好處?
    • 它加快了網路連線速度
    • 它壓縮文檔存儲
    • 它改變了模型的標記器
    • 它過濾掉檢索到的實際上對查詢沒有幫助的段落
  6. How can reflection-token scores affect Self-RAG inference without a new training run?
    • Every query must update all model parameters
    • Weighted critique scores can be combined with generation likelihood during segment-level candidate search
    • The critique scores must replace all source text with random tokens
    • Changing a score weight guarantees all factual errors disappear
  7. What should an evaluator infer from a high Self-RAG IsSupported probability alone?
    • The model predicts support, which still needs checking against the actual evidence
    • The answer has been independently verified as true
    • No source passage needs to be read
    • Every external document is guaranteed accurate
  8. At what granularity can Self-RAG adapt retrieval while developing a response?
    • Only when the operating system restarts
    • Only after the final answer has been published
    • As generated segments progress, rather than only one fixed retrieval at the beginning
    • Only by changing the tokenizer's vocabulary permanently