引導式測驗 · 中等 等級
Kahneman-Tversky 優化測驗
檢查 KTO 的二元合意性回饋、前景理論動機、參考點獎勵以及跨設定的限制。
問題 1 的 8
KTO 需要什麼樣的數據?
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。
- KTO 需要什麼樣的數據?
- 各個輸出上的二進位所需/不需要的標籤
- 對選定和拒絕的答案對進行排名
- 僅未標記的文本
- 數字獎勵分數從 0 到 100
- KTO 的靈感來自於哪些作品?
- 愛因斯坦的相對論
- 卡尼曼和特沃斯基的前景理論
- 達爾文的進化論
- 香農資訊理論
- KTO 中使用的「損失厭惡」是什麼?
- 完全避免任何損失函數
- 拒絕接受損失訓練
- 懲罰不良產出比獎勵理想產出更嚴厲
- 使學習率為負值
- What kind of feedback can KTO learn from without requiring paired comparisons?
- Only a ranked list of five outputs
- Only a reference translation
- Only a token-level grammar mask
- A binary label indicating whether an individual output is desirable
- 在 KTO 中,偏差是相對於什麼來測量的?
- 參考點或基線獎勵水平
- 參數總數
- GPU溫度
- 詞彙量大小
- What does the KTO reference point help define?
- The number of transformer layers
- Whether an output’s implicit reward is above or below a baseline
- A fixed translation in another language
- The model’s maximum context length
- Does KTO’s use of binary desirability feedback prove it will outperform preference-pair methods?
- Yes, binary feedback is always superior
- Yes, it eliminates the need to train a model
- No; performance depends on the model, data, and evaluation setting
- No, because binary feedback cannot be used for language models
- Which behavioral-economics framework motivates KTO?
- Classical signal detection theory
- Bayesian network structure learning
- The efficient-market hypothesis
- Kahneman and Tversky’s prospect theory