引导式测验 · 中等 等级
温度和采样测验
测试 Logit 温度缩放、贪婪解码、top-k 和核心候选过滤器、数值截止以及可预测生成的限制。
问题 1 的 8
升高温度对模型的输出有何影响？
预览每个问题8 问题
当您参加测验时，就会出现答案和解释。
- 升高温度对模型的输出有何影响？
- 使其更具确定性和重复性
- 永久更改模型的权重
- 禁用softmax函数
- 使其更加随机和多样化，有更多机会出现不寻常的标记
- 在采样之前，模型的原始标记分数 (logits) 通过以下方式转换为概率：
- top-k 过滤器
- Softmax函数
- 反向传播
- 分词器
- What determines the candidate set in top-p sampling, compared with top-k?
- Top-p always keeps exactly ten tokens
- Both methods always keep every vocabulary token
- Top-p uses the smallest leading set reaching a cumulative probability threshold; top-k uses a count
- Top-k uses a cumulative threshold while top-p uses a fixed count
- 从力学上来说，温度实际上对逻辑有什么影响？
- 在softmax之前将每个logit除以T
- 为每个 logit 添加一个随机数
- 删除最低的 logits
- 将最终概率乘以 T
- For a structured-data task, what does choosing more predictable token selection fail to establish?
- That a decoding setting was chosen
- That the generated content is correct or valid
- That token scores can be compared
- That a model can produce an output
- How does greedy decoding choose the next token?
- It literally divides every logit by zero
- It uniformly samples every vocabulary token
- It permanently retrains the model
- It selects a highest-scoring eligible token rather than sampling
- Two token logits differ by six before temperature scaling. With positive temperature two, what is their scaled difference?
- 三
- Twelve
- 六
- 零
- For probabilities 0.55, 0.30, 0.10, and 0.05, how many leading tokens does top-p 0.80 keep?
- 一
- 三
- 两个
- 四