Metodebevisregister

Direct Preference Optimization

Direct Preference Optimization (DPO) is a training method introduced by Rafailov et al. in 2023 that aligns language models with human preferences without requiring an explicit reward model. By directly optimizing for preference pairs (better response vs worse response), DPO simplifies the training pipeline compared to reinforcement learning from human feedback (RLHF).

Sources recorded, not reviewed

Kilderegister

Siteringer kopiert ordrett fra metodens kilderegister. Ingen påstandsnivåverifisering er underforstått fra dem.

Direct Preference Optimization: Your Language Model is Secretly a Reward Model

Taksonomisk metoderegister · ml-model / deep-learning

Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., & Finn, C. (2023). Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290. · URL

Åpne full metode

Kuraterte påstander

Påstander lagret i bevishovedboken, hver med sin egen vurdering.

Ingen kuraterte påstander ennå

Denne visningen finner ikke opp en påstandsvurdering når hovedboken ikke har noen.

Relaterte metoder

Generert fra metodegrafen og vist som maskinforslåtte relasjoner – ingen bevispåstand er underforstått.

Same method familyLatent Diffusion Modelsmachine-suggested · Relational suggestion, not evidence.Same method familyMamba (State Space Model)machine-suggested · Relational suggestion, not evidence.Same method familyMasked Autoencodersmachine-suggested · Relational suggestion, not evidence.Same method familyQLoRAmachine-suggested · Relational suggestion, not evidence.