base model

A model that has finished pretraining but has not yet been aligned.

base model = pretrained, not yet instruction-tuned or preference-aligned

A base model is the raw output of pretraining: it has read enormous amounts of text and absorbed grammar, facts, and patterns, but it has had no instruction tuning and no preference alignment. So it only does what pretraining trained it to do — continue the text it is given — rather than treat your input as a request and answer it. Turning a base model into the assistant you chat with is exactly the job of post-training; alignment is the difference between the two.

Appears in

Likely is not good Post-training · pt 1
Show it good answers Post-training · pt 2
Learn a critic Post-training · pt 4

Nearby in the brain

instruction tuning model alignment post-training preference alignment LoRA Bradley-Terry model preference data reward