hendel-llama-7b-architecture

IN premise — summaries/2026/08/24/hendel-2023-icl-task-vectors-sR-references.md

Created 2026-08-25T02:58:04+00:00

LLaMA 7B has 4096 hidden dimensions, 32 layers, and 32 attention heads.

Summary

This is the structural blueprint of the LLaMA 7B model: how wide its internal representations are, how many sequential stages of processing it runs, and how many parallel "spotlight" mechanisms it uses to track relationships between words. These three numbers set the baseline for everything else — they determine the model's capacity, its memory and compute footprint, and how it compares to other architectures in the system.