Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive.
Arka PalDeep KarkhanisSamuel DooleyManley RobertsSiddartha NaiduColin WhitePublished in: CoRR (2024)
Keyphrases
- failure modes
- positive and negative
- fault tree
- genetic algorithm
- preference elicitation
- individual preferences
- user preferences
- positive feedback
- optimisation algorithm
- machine learning
- feature selection
- multiscale
- search algorithm
- data mining
- multi attribute
- soft constraints
- databases
- positively correlated
- data sets