Temperature and sampling parameters in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Effect of Temperature on Model Output Diversity

When using a language model with a temperature parameter, what happens to the model's output as the temperature increases from 0.1 to 1.0?

A The output becomes more random and diverse, allowing more creative responses.

B The output becomes more deterministic and repetitive, reducing creativity.

C The model stops generating any output because the temperature is too high.

D The output length increases automatically with higher temperature.

💡 Hint

Think about how temperature controls randomness in choosing words.

Higher temperature values increase randomness in the model's word choices, making outputs more diverse and creative. Lower temperatures make the model pick the most likely words, producing more predictable text.

❓ Predict Output

intermediate

Sampling with Top-k Parameter

Given a language model sampling with top-k=3, which option best describes the effect on word selection?

A The model samples from all words with probability above 0.3.

B The model always picks the single most probable word.

C The model only considers the top 3 most probable next words for sampling.

D The model ignores the top 3 words and samples from the rest.

💡 Hint

Top-k limits the candidate words to a fixed number.

Top-k sampling restricts the model to choose the next word only from the top k most likely words, increasing output diversity while avoiding very unlikely words.

❓ Hyperparameter

advanced

Choosing Temperature for Balanced Creativity

You want your language model to generate responses that are creative but still coherent. Which temperature value is most suitable?

A 3.0

B 0.5

C 1.5

D 0.01

💡 Hint

Very low temperatures make output boring; very high make it chaotic.

A temperature around 0.5 balances randomness and coherence, producing creative yet sensible outputs. Very low values produce repetitive text; very high values produce nonsensical text.

❓ Metrics

advanced

Impact of Sampling Parameters on Perplexity

How does increasing temperature during sampling affect the perplexity metric measured on generated text?

A Perplexity generally increases because outputs become less predictable.

B Perplexity decreases because outputs become more accurate.

C Perplexity remains unchanged as it depends only on training data.

D Perplexity becomes zero because the model guesses perfectly.

💡 Hint

Higher randomness means less predictable text.

Higher temperature increases randomness, making generated text less predictable and thus increasing perplexity, which measures uncertainty in predictions.

🔧 Debug

expert

Unexpected Output with Temperature and Top-p Sampling

Consider this code snippet using a language model with temperature=0.8 and top-p=0.9. The output is unexpectedly repetitive and lacks diversity. Which is the most likely cause?

Prompt Engineering / GenAI

model.generate(prompt, temperature=0.8, top_p=0.9, top_k=0)

A The model is ignoring temperature when top-p is set.

B Temperature 0.8 is too low to allow any randomness.

C Top-k is set to 0, which disables top-k filtering, causing less diversity.

D Top-p 0.9 is too restrictive, limiting word choices too much.

💡 Hint

Top-p controls cumulative probability cutoff for candidate words.

Top-p=0.9 means sampling only from the smallest set of words whose probabilities sum to 90%. If this set is very small, it limits diversity despite temperature. Top-k=0 disables top-k filtering, so it does not reduce diversity.