引導式測驗 · 中等 等級
上下文視窗測驗
檢查令牌上下文預算、輸入和輸出分配、長輸入評估限制、檢索權衡以及請求超出容量時的行為。
問題 1 的 8
模型的上下文視窗衡量什麼?
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。
- 模型的上下文視窗衡量什麼?
- 聊天介面的螢幕分辨率
- 模型可以一次考慮的最大文字量(以標記為單位)
- 多少用戶可以同時聊天
- 模型的總訓練資料集大小
- How should a token be interpreted when budgeting language-model context?
- As exactly one full word in every language
- As exactly one sentence in every tokenizer
- As a tokenizer-defined unit whose relation to words depends on the model and text
- As the API password used to send a request
- In a decoder-style request with a shared input-and-continuation budget, which supplied material uses that budget?
- Instructions, conversation turns, document tokens, and the generated continuation
- Only the latest user message
- Only stored application files that were never supplied
- Only generated output, with input always free
- Under a dense full-attention pairwise-interaction model, how does comparison count scale when a request's token length doubles?
- It becomes only twice as large under that model
- It stays unchanged regardless of length
- It becomes half as large
- It becomes four times as large under the quadratic model
- 「迷失在中間」問題是指模型傾向:
- 當輸入很長時崩潰
- 將文件的中間部分翻譯成另一種語言
- 長輸入的開頭和結尾的資訊比中間的資訊更好回憶
- 完全忽略第一句和最後一句
- Why can retrieval be combined with a language model's limited context window?
- It guarantees the lowest total cost for every workflow
- It selects passages to supply rather than requiring the whole knowledge base as input
- It permanently increases the model's supported token capacity
- It removes the need to evaluate generated answers
- In an illustrative shared context budget of 8,000 tokens, input uses 6,500. How many tokens remain before other limits or overhead?
- 6,500
- 8,000
- 14,500
- 1,500
- What can happen when an input exceeds a language model interface's supported context?
- Rejection, truncation, or application-side shortening, depending on the setup
- Automatic permanent memory of all excluded text
- Guaranteed silent deletion of the oldest message in every system
- Automatic expansion of every model's trained architecture