SAE-TS Effects Dataset
This dataset contains pre-computed feature effects for the SAE-TS (SAE-Targeted Steering) method described in our paper Improving Steering Vectors by Targeting Sparse Autoencoder Features.
Contents
The dataset contains two files:
effects_2b.pt: Pre-computed effects for Gemma-2B modeleffects_9b.pt: Pre-computed effects for Gemma-9B model
Each file is a PyTorch saved dictionary containing:
features: The steering vectors used (shape: [num_features, d_model])effects: The measured effects of each steering vector (shape: [num_features, d_sae])
Usage
These files are used by the SAE-TS training script. To use them:
import torch
from huggingface_hub import hf_hub_download
# Download the data
path = hf_hub_download(repo_id="schalnev/sae-ts-effects",
filename="effects_2b.pt") # or effects_9b.pt
# Load the data
data = torch.load(path)
features = data['features']
effects = data['effects']
License
This dataset is released under the MIT license. See LICENSE for more information.