Complete extraction of all individual concepts from every Sparse Autoencoder (SAE) released on Neuronpedia, plus all public features from Anthropic's Towards Monosemanticity (2023) and Scaling Monosemanticity (2024) papers.
from datasets import load_dataset
Full Neuronpedia dataset (77M rows, streaming recommended)
ds = load_dataset("hbe/neuronpedia-sae-concepts", split="train", streaming=True)
