Instructions to use suchitg/sae-compression-gemma-2-2b-pruned-sae-openwebtext-0.5 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- SAELens
How to use suchitg/sae-compression-gemma-2-2b-pruned-sae-openwebtext-0.5 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
SAEs for use with the SAELens library
This repository contains the following SAEs:
- blocks.0.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.1.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.2.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.3.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.4.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.5.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.6.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.7.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.8.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.9.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.10.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.11.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.12.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.13.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.14.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.15.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.16.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.17.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.18.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.19.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.20.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.21.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.22.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.23.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.24.(hook_resid_post, hook_mlp_out, attn.hook_z)
- blocks.25.(hook_resid_post, hook_mlp_out, attn.hook_z)
Load these SAEs using SAELens as below:
from sae_lens import SAE
sae, cfg_dict, sparsity = SAE.from_pretrained("suchitg/sae-compression-gemma-2-2b-pruned-sae-openwebtext-0.5", "<sae_id>")
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support