Slide 15
Slide 15 text
Automated Promptingの例
SAMMO(Structure-aware Multi-Objective Metaprompt
Optimization)
d プロンプトエンジニアリングと最適化のためのPythonライブラリ
Instruction Optimizationの3つのステッf
dw Defining the set of initial candidateh
qw 初期プロンプト、データセット、設o
vw Define mutation operatorh
qw e.g.) 既存の指示を単に言い換えi
w Run beam search
"examples": [
{
"input": "Speaker 1: 'But aren't you afraid?' Speaker 2: 'Ma'am, sharks never
attack anybody.'",
"target_scores": {
"yes": 0.0,
"no": 1.0
}
},
{
"input": "Speaker 1: 'Do you want to quit?' Speaker 2: 'I've never been the
type of person who throws in the towel when things get tough.'",
"target_scores": {
"yes": 0.0,
"no": 1.0
}
}
]
https://github.com/microsoft/sammo