Open Models Almanac
Methodsmodel architecturepositional encoding

single source · model architecture

Learned input-dependent relative-position bias

A learned bias depending on the input supplies relative-position information instead of RoPE; one implementation limits it to the preceding 1,024 tokens.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See positional encoding for the methods filed beside it.

Filed under model architecture › positional encoding