What is min-p sampling?

hard

Answer

$\mathrm{probability}\; < \;\operatorname{min}_{p}\; \cdot \;P(\mathrm{top}_{\mathrm{token}})$
$\operatorname{min}_{p} = 0.05$
$Adaptive to confidence: when the model is very confident, only the top few tokens survive; when it's uncertain, many pass through.$
$Empirically produces higher-quality creative writing than top-p at similar diversity.$
$Adopted in llama.cpp, vLLM, HuggingFace.$

Check yourself — multiple choice

Fixed threshold on probability
Keep tokens with $\mathrm{prob}\; \ge \;\operatorname{min}_{p}$ × prob(top token) — adaptive relative filter, better creative writing than top-p
Same as top-p
Removes sampling

min-p: relative-to-top-token threshold → adaptive, high-quality alternative to top-p.

#generation#decoding#sampling

214 interview questions in this topic.