Update README.md

README.md

@@ -11,17 +11,20 @@ These SAEs were trained with [SAE Lens](https://github.com/jbloomAus/SAELens) an
 
 All training hyperparameters are specified in cfg.json.
 
-They are loadable using SAE via a few methods. A method that currently works (but may be replaced shortly by a more convenient method) would be the following:
+They are loadable using SAE via a few methods. The preferred method is to use the following:
 
 ```python
 import torch
-from sae_lens.training.session_loader import LMSparseAutoencoderSessionloader
+from transformer_lens import HookedTransformer
+from sae_lens import SparseAutoencoder, ActivationsStore
 
 torch.set_grad_enabled(False)
-path = "path/to/folder_containing_cfgjson_and_safetensors_file"
-model, sae, activation_store = LMSparseAutoencoderSessionloader.load_pretrained_sae(
- path, device = "cuda",
+model = HookedTransformer.from_pretrained("gemma-2b")
+sparse_autoencoder = SparseAutoencoder.from_pretrained(
+ "gemma-2b-res-jb", # to see the list of available releases, go to: https://github.com/jbloomAus/SAELens/blob/main/sae_lens/pretrained_saes.yaml
+ "blocks.0.hook_resid_post" # change this to another specific SAE ID in the release if desired. 
 )
+activation_store = ActivationsStore.from_config(model, sparse_autoencoder.cfg)
 ```
 
 ## Resid Post 0