dhruv-anand-aintech/llm-autoencoder-interpretability-sae-qwen-qwen3-4b-layer18

Sparse autoencoder trained in Modal on Qwen/Qwen3-4B layer 18.

Usage

from safetensors.torch import load_file
import json
import torch
from src.sae import SAEConfig, SparseAutoencoder

cfg = SAEConfig(**json.load(open("config.json")))
sae = SparseAutoencoder(cfg)
sae.load_state_dict(load_file("model.safetensors"))
sae.eval()