Instructions, sampling, and verification
Compare generation settings without confusing consistency with factual accuracy.
- Compare generation settings without confusing consistency with factual accuracy.
System instructions guide an assistant but do not guarantee obedience or secure tool access. The application must enforce permissions. Temperature rescales token scores before softmax: lower positive values concentrate probability, higher values flatten it. Top-k keeps a fixed number of candidates; top-p keeps a set reaching a chosen cumulative probability. Neither setting checks facts. Greedy decoding chooses a highest-scoring token rather than sampling.
A closer look
Try greedy and sampling modes several times in the builder. Sampling can repeat the same result. Temperature zero does not guarantee identical output across real systems: model versions, context, tools, and numerical execution also matter. Then compare a vague question with one giving a goal, context, and a way to check the answer. Test a factual claim against an original source; use hints and explain the result in your own words.
Explain it yourself
Can a low-temperature model confidently repeat a false answer?
Compare with an explanation
Yes. Concentrated probabilities are not evidence of truth. Check the claim independently.
You mark this yourself. Nothing is graded here.