Min-p Sampling (Minimum Probability Threshold・2023年-2024年)（ミンピーサンプリング）
2023年Nguyen et al.発表Min-p Sampling paper・Industry-emerging Minimum Probability threshold sampling paradigm + Industry-leading min_p=0.05-0.1 default + Industry-emerging Top-p alternative paradigm + Industry-leading creative writing + low-temperature scenarios advantage + Industry-leading Hugging Face + llama.cpp Industry-emerging adoption。
概要
Min-p Sampling は、2023年Nguyen et al.発表Min-p Sampling paper・Industry-emerging Minimum Probability threshold sampling paradigm 2023年-2024年 + Industry-emerging Top-p alternative paradigm Pioneer position確立。Min-p Sampling specifications = Industry-emerging Minimum Probability threshold sampling paradigm (Industry-emerging Minimum Probability threshold + Industry-leading min_p threshold absolute filter + Industry-leading Top-p cumulative vs Min-p absolute paradigm差異) + Industry-leading min_p=0.05-0.1 default (Industry-leading min_p=0.05-0.1 default + Industry-leading conservative absolute probability filter) + Industry-emerging Top-p alternative paradigm (Industry-leading Top-p cumulative probability vs Min-p absolute probability paradigm + Industry-emerging Top-p alternative paradigm Pioneer) + Industry-leading creative writing + low-temperature scenarios advantage (Industry-leading creative writing scenarios + Industry-leading low-temperature scenarios + Industry-emerging Min-p Industry-leading creative writing advantage) + Industry-leading Hugging Face + llama.cpp Industry-emerging adoption (Industry-leading Hugging Face Transformers Industry-emerging Min-p support + Industry-leading llama.cpp Industry-leading Min-p adoption + Industry-leading Open-source community Industry-emerging Min-p adoption)。
主な特徴・仕組み
- Authoring: Nguyen et al. 2023年発表
- Paradigm: Industry-emerging Minimum Probability threshold sampling
- Threshold: Industry-leading min_p=0.05-0.1 default
- Filter Type: Industry-leading absolute probability filter
- Industry Position: Industry-emerging Top-p alternative Pioneer
- Advantage: Industry-leading creative writing + low-temperature scenarios
- Adoption: Industry-leading Hugging Face + llama.cpp Industry-emerging adoption
- vs Top-p: Industry-leading absolute vs cumulative probability paradigm差異
- Industry-Emerging: Industry-emerging 2023年Min-p Pioneer paradigm
- Use Cases: Industry-leading creative writing + low-temperature LLM scenarios
スペック比較表
| LLM Sampling Method (2024) | Approach | Default Value | Industry Position |
|---|---|---|---|
| Min-p Sampling | Minimum absolute probability threshold | min_p=0.05-0.1 | Industry-emerging Min-p Pioneer |
| Top-p Nucleus Sampling | Cumulative probability threshold | p=0.9 default | Industry-leading dominant sampling |
| Top-k Sampling | Fixed top-k tokens | k=40 default | Industry-leading classic Pioneer |
| Temperature Sampling | Probability scaling | T=0.7-1.0 default | Industry-leading 全LLM dominant standard |
| Top-nsigma Sampling | Statistical sigma threshold | nsigma=1.0 | Industry-emerging 2024年Pioneer |
具体例・対応製品
- Min-p Sampling Paper (Nguyen et al.・2023年): Industry-emerging Pioneer paper
- Industry-leading Hugging Face Transformers Min-p support: Industry-emerging framework adoption
- Industry-leading llama.cpp Min-p adoption: Industry-leading lightweight framework adoption
- Industry-leading min_p=0.05-0.1 default: Industry-leading absolute probability filter
- Industry-leading creative writing + low-temperature scenarios: Industry-leading use case
- 競合 Top-p + Top-k + Temperature: Industry-leading sampling competitors
自作PCでの選び方・注意点
Min-p Sampling は「Industry-emerging Minimum Probability threshold sampling paradigm + Industry-leading min_p=0.05-0.1 default」「Industry-emerging Top-p alternative paradigm + Industry-leading creative writing + low-temperature scenarios advantage」用途のIndustry-emerging Min-p sampling 2023年Nguyen et al. product。Industry-emerging Minimum Probability + min_p=0.05-0.1 default (Industry-leading min_p absolute probability filter vs Industry-leading Top-p cumulative probability filter + Industry-leading creative writing scenarios + low-temperature scenarios advantage) で Industry-emerging absolute probability paradigm + Industry-leading min_p=0.05-0.1 default + Industry-leading creative writing advantage。Industry-emerging Top-p alternative paradigm (Industry-leading Top-p cumulative probability vs Min-p absolute probability paradigm差異 + Industry-emerging Top-p alternative Pioneer) で Industry-emerging Top-p alternative paradigm + Industry-leading cumulative vs absolute paradigm差異 + Industry-emerging Min-p Pioneer position。Industry-leading Hugging Face + llama.cpp Industry-emerging adoption (Industry-leading Hugging Face Transformers Industry-emerging Min-p support + Industry-leading llama.cpp Industry-leading Min-p adoption + Industry-leading Open-source community Industry-emerging Min-p adoption) で Industry-leading Hugging Face + llama.cpp Industry-emerging Min-p adoption + Industry-leading Open-source community adoption。但しIndustry-leading Top-p 2020年dominant vs Min-p 2023年emerging paradigm trade-off (Industry-leading Top-p mature dominant adoption vs Industry-emerging Min-p 2023年emerging paradigm + Industry-leading Min-p creative writing + low-temperature scenarios specific advantage trade-off) で Industry-leading Top-p mature dominant vs Industry-emerging Min-p emerging trade-off + Industry-leading creative writing + low-temperature scenarios specific advantage判断要。
関連用語との違い
- vs Top-p Nucleus Sampling (cumulative): Min-pはAbsolute probability + min_p=0.05-0.1 + 2023年Nguyen Pioneer・Top-pはCumulative probability + p=0.9 + 2020年Holtzman dominant
- vs Top-k + Temperature: Min-pはAbsolute probability threshold + Top-p alternative・Top-kはFixed size + TemperatureはProbability scaling
- vs Top-nsigma Sampling (2024年): Min-pは2023年Nguyen Pioneer・Top-nsigmaは2024年Statistical sigma Pioneer
よくある質問（FAQ）
Q1: Min-p vs Top-p 違いは？ A: Min-p (Industry-emerging absolute probability threshold + min_p=0.05-0.1 + creative writing + low-temperature scenarios advantage + Nguyen 2023年) vs Top-p (Industry-leading cumulative probability + p=0.9 + dominant adoption + Holtzman 2020年)・Industry-leading creative + low-temperature = Min-p + Industry-leading mature dominant = Top-p preference judgment。
Q2: Industry-leading min_p=0.05-0.1 default value は？ A: Industry-leading min_p=0.05-0.1 default (Industry-leading absolute probability filter + Industry-leading conservative threshold + Industry-emerging Min-p Industry-emerging paradigm default)。
Q3: Industry-leading creative writing + low-temperature scenarios advantage は？ A: Industry-leading creative writing + low-temperature advantage (Industry-leading creative writing scenarios + Industry-leading low-temperature scenarios + Industry-emerging Min-p Industry-leading creative writing advantage + Industry-leading absolute probability filter advantage in creative writing)。
まとめ
Min-p Sampling = 2023年Nguyen et al.発表のMinimum Probability threshold sampling paradigm。Industry-emerging absolute probability threshold + Industry-leading min_p=0.05-0.1 default + Industry-emerging Top-p alternative paradigm + Industry-leading creative writing + low-temperature scenarios advantage + Industry-leading Hugging Face Transformers + llama.cpp Industry-emerging adoption + Industry-emerging Min-p sampling paradigm Pioneer position確立。