引導式測驗 · 中等 等級
帶約束的引導束搜尋測驗
涵蓋束搜尋、詞彙和語法約束、動態束分配以及語法和語義有效性之間的區別。
問題 1 的 8
普通波束搜尋在每個生成步驟中保留什麼?
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。
- 普通波束搜尋在每個生成步驟中保留什麼?
- Top-k 最可能的部分序列
- 僅單一最可能的標記
- 所有可能的序列
- 序列的隨機樣本
- 像網格波束搜尋這樣的詞法約束方法如何組織波束?
- 僅按輸出長度
- 每個梁滿足多少個所需的約束
- 按字母順序
- 僅靠模型的置信度
- 在語法約束解碼中,如何防止無效輸出?
- 透過在生成後刪除不良令牌
- 透過有效範例重新訓練模型
- 透過屏蔽標記機率,只允許使用語法有效的標記
- 透過降低溫度
- 動態光束分配解決什麼問題?
- 它消除了約束的需要
- 它使模型更大
- 它完全消除了波束搜索
- 它有效地跨約束進度等級共享梁槽
- When can grammar-constrained decoding produce syntactically valid JSON?
- When the grammar constraint is enforced through completion and the output is not truncated
- Whenever JSON is mentioned in the prompt
- Even if the runtime stops before required closing tokens
- Regardless of whether the active grammar matches the intended schema
- 為什麼硬約束會在沒有特殊處理的情況下擠出好的序列?
- 它們使所有光束都相同
- 忽略約束的高機率光束可能會主導 top-k
- 它們會減慢 GPU 速度
- 他們增加詞彙量
- A lexical constraint requires a translation to include “bridge” and “river.” What does constrained beam search enforce?
- The output has exactly two words
- The model uses only one beam
- The completed output includes the required terms
- The translation is factually accurate
- A grammar-constrained generator returns JSON that parses successfully. What does that not establish by itself?
- That the output follows the grammar
- That required fields can be present
- That the syntax is valid
- That the values are factually correct or appropriate for the application