Nucleus 和 Top-k 采样测验

当您参加测验时，就会出现答案和解释。

1. top-k 采样如何限制 token 选择？
   • 它只保留 k 个最高概率的标记，然后在其中进行采样
   • 它保留令牌直到累积概率超过 k
   • 它总是选择单个最有可能的标记
   • 它仅删除 k 个最不可能的标记
2. 什么定义了核（top-p）采样中的“核”？
   • 训练数据中出现频率最高的 k 个单词
   • 概率总和至少为 p 的最小标记集
   • 单个最高概率的令牌
   • 所有概率高于 50% 的代币
3. top-p 相对于 top-k 的主要优势是什么？
   • 计算总是更快
   • 它从不产生随机输出
   • 候选标记的数量适应模型的置信度
   • 它忽略了令牌概率
4. 采样前温度参数有什么作用？
   • 它设置最大序列长度
   • 它选择输出的语言
   • 它计算词汇量的大小
   • 它将概率分布重新调整为更尖锐或更平坦
5. What result did the nucleus-sampling paper report for its tested open-ended generation tasks?
   • It improved diversity without sacrificing fluency and coherence in the reported comparisons
   • It guaranteed factual correctness in every output
   • It eliminated all sampling randomness
   • It improved translation quality on every benchmark
6. For a greedy decoding baseline with no token sampling, how should generation be configured?
   • Disable sampling and use greedy decoding to select the highest-probability token at each step
   • Use high temperature and a large nucleus threshold
   • Randomize the temperature at every step
   • Keep sampling enabled but increase top-p
7. After top-k truncation removes tokens, what happens to the remaining sampling probabilities?
   • They remain assigned their original total mass below one
   • They are replaced with equal probabilities
   • They are renormalized over the retained top-k tokens
   • They are converted into beam-search scores
8. What did Holtzman et al. report about likelihood-maximizing decoding for open-ended text?
   • It always yields the most diverse human-like passages
   • 它消除了对语言模型的需要
   • It guarantees factual correctness
   • It can produce bland and repetitive text despite strong likelihood scores