Aquinlabs/llama-3.2-1b-sae

Llama 3.2 1B SAEs (layers 0–15)

Sparse autoencoders for meta-llama/Llama-3.2-1B-Instruct.

Layout

layer0_32768_top32/
  sae.pt
  meta.json
...
layer15_32768_top32/
  sae.pt
  meta.json
umap_points.json

Specs

Field	Value
Base model	meta-llama/Llama-3.2-1B-Instruct
Layers	0–15 (16 SAEs)
d_model	2048
d_sae	32768
Activation	TopK (k=32)
Architecture	topk
Package size	~8.1 GiB

Files

• sae.pt — SAE checkpoint for that layer
• meta.json — layer metadata (model_slug, layer, d_model, d_sae, k, arch)
• umap_points.json — feature UMAP coordinates

License

• Base model: Llama 3.2 Community License (Meta). Using these SAEs with Llama 3.2 remains subject to that license.
• Accept Meta’s license on Hugging Face before downloading/using the base model.

Acknowledgements

• Base model: meta-llama/Llama-3.2-1B-Instruct