引導式測驗 · 中等 等級

上下文擴展測驗的位置插值

測試 RoPE 位置壓縮、外推、擴展因子、頻率感知方法、模型自適應和長上下文資源成本。

問題 1 的 8

位置插值的核心思想是什么?

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. 位置插值的核心思想是什么?
    • 添加更多關注頭
    • 重新調整(縮小)位置索引,使它們保持在訓練範圍內,而不是推斷
    • 完全刪除位置嵌入
    • 在长序列上从头开始重新训练模型
  2. Which positional scheme did the original Position Interpolation work extend through position-index rescaling?
    • A fixed output-vocabulary lookup table only
    • No positional information of any kind
    • Rotary positional embeddings, or RoPE
    • An image compression format rather than model positions
  3. Why can unmodified positional extrapolation degrade a RoPE model on much longer inputs?
    • The sequence introduces positional and attention conditions beyond the model's training experience
    • A longer context automatically retrains every parameter
    • RoPE removes all attention from the model
    • The tokenizer necessarily loses its vocabulary at a specific length
  4. 如果將上下文長度從 L 擴展到 L',基本 PI 將什麼重新縮放因子應用於位置 m?
    • m 乘以 L'/L
    • 將 L' 加到 m
    • 將 m 設定為零
    • m 除以 L'/L(即 m / (L'/L))
  5. How does YaRN's frequency-aware approach differ from uniform Position Interpolation?
    • It removes every query and key from attention
    • It changes only the output vocabulary size
    • It treats frequency bands differently to preserve local detail while extending positional range
    • It guarantees that all long-context tasks have identical accuracy
  6. What makes the original Position Interpolation workflow different from pretraining a new long-context model from scratch?
    • It eliminates every need for computation
    • It adapts an existing pretrained model toward the tested longer window with rescaling and fine-tuning
    • It requires replacing all learned parameters with random values
    • It proves all models can extend indefinitely without evaluation
  7. Does basic Position Interpolation inherently add new attention heads or enlarge the trainable architecture?
    • Yes; every doubled context requires twice as many attention heads
    • Yes; the vocabulary must expand by the context-length factor
    • Only the model's filename changes, with no effect on positions
    • No; it changes the position coordinates used by the existing model
  8. Why must a longer-context PI deployment still assess memory and computation requirements?
    • It processes more tokens and stores relevant state; compressing position coordinates does not remove that work
    • The number of actual input tokens shrinks automatically with their positions
    • All KV-cache storage becomes zero after rescaling
    • Every extended model has unchanged latency regardless of input length