A brain atlas for meta-llama/Llama-3.2-3B-Instruct, the 3B instruction-tuned member of the Llama 3.2 family. This is not a chat dataset or a benchmark - it is an internal-mechanics map built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.
If you want to know where an instruction-tuned model keeps its register machinery, which directions survive a causal test… See the full description on the dataset page:
https://huggingface.co/datasets/juiceb0xc0de/llama3.2-3b-instruct-atlas.
