Temperature Sampling (Probability Scaling・1985年-2024年)（テンパレチャーサンプリング）
1985年Statistical Mechanics Temperature paradigm起源・1992年Hinton Boltzmann machines応用・Industry-leading Probability scaling sampling paradigm + Industry-leading T=0.7-1.0 default + Industry-leading temperature control + Industry-leading LLM sampling Industry-standard 全LLM Industry-leading adoption + Industry-leading Top-k/Top-p併用 Industry-standard。
概要
Temperature Sampling は、1985年Statistical Mechanics Temperature paradigm起源・1992年Hinton Boltzmann machines応用・Industry-leading Probability scaling sampling paradigm 1985年-2024年 + Industry-leading 全LLM Industry-leading dominant adoption Industry-standard position確立。Temperature Sampling specifications = Industry-leading Probability scaling sampling paradigm (Industry-leading Probability scaling sampling + Industry-leading T scalar temperature parameter + Industry-leading softmax temperature scaling paradigm) + Industry-leading T=0.7-1.0 default Industry-standard (Industry-leading T=0.7 conservative + T=1.0 default + T=1.5 creative + Industry-leading LLM Industry-standard temperature control) + Industry-leading temperature control (Industry-leading T低い conservative deterministic + Industry-leading T高い creative random + Industry-leading temperature control creativity vs determinism paradigm) + Industry-leading LLM sampling Industry-standard 全LLM adoption (Industry-leading 全LLM model + 全LLM framework + 全LLM API Industry-leading Temperature sampling Industry-standard adoption) + Industry-leading Top-k/Top-p併用 Industry-standard (Industry-leading Temperature + Top-k or Temperature + Top-p Industry-leading combination Industry-standard adoption paradigm)・Industry-leading Hinton Industry-leading 1992年Boltzmann machines paper Pioneer credibility backing。
主な特徴・仕組み
- Origin: 1985年Statistical Mechanics + 1992年Hinton Boltzmann machines応用
- Paradigm: Industry-leading Probability scaling sampling
- Parameter: Industry-leading T scalar temperature parameter
- Default: Industry-leading T=0.7-1.0 default Industry-standard
- Industry Position: Industry-leading 全LLM dominant adoption Industry-standard
- Control: Industry-leading creativity vs determinism control paradigm
- Hinton Foundation: Industry-leading 1992年Boltzmann machines Pioneer paradigm
- Combination: Industry-leading Top-k/Top-p併用 Industry-standard
- Heritage: Industry-leading 40年Statistical Mechanics + LLM application
- Use Cases: Industry-leading 全LLM creativity vs deterministic control
スペック比較表
| LLM Sampling Method (2024) | Approach | Default Value | Industry Position |
|---|---|---|---|
| Temperature Sampling | Probability scaling | T=0.7-1.0 default | Industry-leading 全LLM dominant standard |
| 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 |
| Min-p Sampling | Minimum probability threshold | min_p=0.05-0.1 | Industry-emerging Min-p paradigm |
| Top-nsigma Sampling | Statistical sigma threshold | nsigma=1.0 | Industry-emerging 2024年Pioneer |
具体例・対応製品
- Temperature Sampling (1985年Statistical Mechanics起源): Industry-leading 40年heritage
- Hinton Industry-leading 1992年Boltzmann machines応用 Pioneer: Industry-leading Pioneer
- Industry-leading 全LLM model + framework + API Industry-standard adoption: Industry-leading universal
- Industry-leading T=0.7-1.0 default Industry-standard: Industry-leading default
- Industry-leading Temperature + Top-k/Top-p combination standard: Industry-leading combination
- 競合 Top-p + Top-k + Min-p + Top-nsigma: Industry-leading sampling competitors
自作PCでの選び方・注意点
Temperature Sampling は「Industry-leading Probability scaling sampling paradigm + Industry-leading T=0.7-1.0 default」「Industry-leading temperature control + Industry-leading LLM sampling Industry-standard 全LLM adoption」「Industry-leading Top-k/Top-p併用 Industry-standard」用途のIndustry-leading 全LLM dominant sampling 1985年Statistical Mechanics + 1992年Hinton応用 paradigm。Industry-leading Probability scaling + T=0.7-1.0 default (Industry-leading T=0.7 conservative + T=1.0 default + T=1.5 creative + Industry-leading LLM Industry-standard temperature control) で Industry-leading T=0.7-1.0 default + Industry-leading conservative + default + creative control。Industry-leading 全LLM Industry-standard adoption (Industry-leading 全LLM model + 全LLM framework + 全LLM API Industry-leading Temperature sampling Industry-standard adoption + Industry-leading Hugging Face + OpenAI + Anthropic + 全LLM API Temperature parameter Industry-standard) で Industry-leading 全LLM universal Industry-standard + Industry-leading Hugging Face + OpenAI + Anthropic universal adoption。Industry-leading Top-k/Top-p併用 Industry-standard (Industry-leading Temperature + Top-k or Temperature + Top-p Industry-leading combination Industry-standard adoption paradigm) で Industry-leading Temperature + Top-k/Top-p combination Industry-standard + Industry-leading LLM Industry-standard combination paradigm。Industry-leading Hinton Industry-leading 1992年Boltzmann machines Pioneer credibility (Industry-leading 2018年Turing Award Hinton + Industry-leading neural network Pioneer + Industry-leading Boltzmann machines Pioneer credibility backing) で Industry-leading Hinton 2018年Turing Award + neural network Pioneer + Boltzmann machines Pioneer credibility backing。但しIndustry-leading Temperature単独 vs Top-k/Top-p併用 standard trade-off (Industry-leading Temperature単独 vs Industry-leading combination standard adoption + Industry-leading Temperature creativity control vs Top-k/Top-p vocabulary control + Industry-leading combination universal standard adoption) で Industry-leading Temperature単独 vs combination standard adoption trade-off + Industry-leading combination universal standard adoption。
関連用語との違い
- vs Top-p + Top-k (vocabulary truncation): TemperatureはProbability scaling paradigm + 通常Top-k/Top-p併用・他はVocabulary truncation paradigms
- vs Greedy Decoding (T=0): TemperatureはT=0.7-1.0 stochastic + Greedy DecodingはT=0 deterministic equivalent
- vs Min-p + Top-nsigma (2024年emerging): Temperatureは1985年-2024年mature universal standard・他は2023-2024年emerging paradigms
よくある質問（FAQ）
Q1: Temperature vs Top-p 違いは？ A: Temperature (Probability scaling T=0.7-1.0 + creativity vs determinism control + Hinton 1992年Pioneer + 通常Top-k/Top-p併用) vs Top-p (Cumulative probability threshold p=0.9 + vocabulary truncation + Holtzman 2020年)・Industry-leading combination Industry-standard = Temperature + Top-p preference judgment。
Q2: Industry-leading T=0.7-1.0 default value は？ A: Industry-leading T=0.7-1.0 default (Industry-leading T=0.7 conservative + T=1.0 default + T=1.5 creative + Industry-leading LLM Industry-standard temperature control + Industry-leading 全LLM API default)。
Q3: Industry-leading Hinton 1992年Boltzmann machines credibility は？ A: Industry-leading Hinton (Industry-leading 2018年Turing Award + Industry-leading neural network Pioneer + Industry-leading 1992年Boltzmann machines Pioneer + Industry-leading Temperature sampling LLM application Pioneer credibility backing)。
まとめ
Temperature Sampling = 1985年Statistical Mechanics起源・1992年Hinton Boltzmann machines応用のProbability scaling sampling paradigm。Industry-leading T=0.7-1.0 default Industry-standard + Industry-leading temperature control creativity vs determinism + Industry-leading 全LLM Industry-standard adoption + Industry-leading Top-k/Top-p併用 Industry-standard combination + Industry-leading Hinton 2018年Turing Award + 1992年Boltzmann machines Pioneer credibility backing + Industry-leading 40年heritage LLM sampling Industry-leading universal standard paradigm position確立。