peal.sparse_dictionaries.ra_sae_decomposition.RASAEDecompositionConfig

class peal.sparse_dictionaries.ra_sae_decomposition.RASAEDecompositionConfig[source]

Bases: SparseDictionaryConfig

Configuration for a pretrained RA-SAE decomposition.

Methods

Attributes

act_size

ckpt_url

config_url

drop_raw_checkpoint_after_fuse

Delete the 4.4 GB raw checkpoint once the fused cache is written.

encoder_model

Backbone whose tokens this dictionary decomposes.

fused_filename

Name of the fused-tensor cache written next to the raw checkpoint.

local_dir

Writable scratch directory.

n_components

repo_id

running_threshold

Activation threshold of the BatchTopK eval branch.

sparse_dictionaries_type

sparse_dictionaries_type: str = 'RASAEDecomposition'
encoder_model: str = 'dino_v2_base_reg'

Backbone whose tokens this dictionary decomposes. Kept for the record and checked against the generator’s encoder tag by consumers that edit z_sem.

repo_id: str = 'matybohacek/RA-SAE-DINOv2-32k'
ckpt_url: str = 'https://huggingface.co/matybohacek/RA-SAE-DINOv2-32k/resolve/main/RA-SAE-DINOv2-32k.pth'
config_url: str = 'https://huggingface.co/matybohacek/RA-SAE-DINOv2-32k/raw/main/config.json'
local_dir: str = '/tmp/ra_sae_weights'

Writable scratch directory. The GPU nodes mount ~/.cache read-only, so this must not default anywhere under it.

fused_filename: str = 'ra_sae_dinov2_32k_fused.pt'

Name of the fused-tensor cache written next to the raw checkpoint.

running_threshold: float | None = None

Activation threshold of the BatchTopK eval branch. None -> read from the published config.json (0.829 for this checkpoint).

drop_raw_checkpoint_after_fuse: bool = False

Delete the 4.4 GB raw checkpoint once the fused cache is written.

n_components: int | None = 32000
act_size: int = 768