Mirostat Perplexity Control Quiz

Answers and explanations appear as you take the quiz.

1. What does Mirostat try to control during generation?
   • A target level of output perplexity or surprise
   • The number of model parameters
   • The training dataset size
   • The GPU temperature
2. How is Mirostat fundamentally different from standard top-k or top-p sampling?
   • It never uses randomness
   • It dynamically adjusts truncation using feedback instead of a fixed cutoff
   • It only works on images
   • It requires retraining the model
3. In Mirostat, what role does the parameter eta play?
   • It sets the vocabulary size
   • It fixes the output length
   • It is the learning rate controlling how fast the threshold adapts
   • It disables sampling
4. What is the main improvement of Mirostat version 2 over version 1?
   • It generates images too
   • It requires more memory
   • It removes the tau parameter
   • It drops the assumption of a Zipfian distribution for simpler, more robust control
5. If Mirostat detects that output is becoming too predictable, it will:
   • Loosen truncation to admit more diverse tokens
   • Tighten truncation to allow fewer tokens
   • Stop generating immediately
   • Increase the model size
6. Which Mirostat settings does current llama.cpp expose?
   • Tokenizer vocabulary and number of layers
   • Mode, target tau, and learning-rate eta
   • Only the prompt length
   • Training epochs and validation loss
7. What does eta control in Mirostat’s feedback update?
   • The size of the model vocabulary
   • The target number of output tokens
   • How strongly the threshold changes in response to the observed error
   • The probability of a correct factual answer
8. What does a Mirostat target not guarantee by itself?
   • That the sampler reads token probabilities
   • That the setting can influence surprise
   • That the output contains tokens
   • That the generated text is factually correct