Instructions to use google/gemma-scope-9b-pt-res with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- SAELens
How to use google/gemma-scope-9b-pt-res with SAELens:
# pip install sae-lens from sae_lens import SAE sae, cfg_dict, sparsity = SAE.from_pretrained( release = "RELEASE_ID", # e.g., "gpt2-small-res-jb". See other options in https://github.com/jbloomAus/SAELens/blob/main/sae_lens/pretrained_saes.yaml sae_id = "SAE_ID", # e.g., "blocks.8.hook_resid_pre". Won't always be a hook point ) - Notebooks
- Google Colab
- Kaggle
Layer 13 saes raising "zipfile.BadZipFile: File is not a zip file"
#5
by MrGonao - opened
I tried both l0 30 and 54 and they get that error when loading the sae.
Same with layer 6 l0 36.
Sorry, I should have been more clear. When downloading these SAEs with path_to_params = hf_hub_download( repo_id="google/gemma-scope-9b-pt-res", filename=f"layer_13/width_131k/average_l0_30"/params.npz", force_download=False, ) and then loading them using params = np.load(path_to_params) leads to the above error for SAEs of width 131k on layer 3 and 6 (at least in the specified l0s). All other layers I tried seem to be working fine.
Oh, yesI
MrGonao changed discussion status to closed