Modelpublic

annasoli/gemma-3-27b-activation-oracle

sourceHugging Faceupdated 9mo agoView on Hugging Face

0likes6downloads

Model Card

LoRA Adapter for SAE Introspection

This is a LoRA (Low-Rank Adaptation) adapter trained for SAE (Sparse Autoencoder) introspection tasks.

Base Model

—Base Model: google/gemma-3-27b-it
—Adapter Type: LoRA
—Task: SAE Feature Introspection

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained("google/gemma-3-27b-it")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-27b-it")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "annasoli/gemma-3-27b-activation-oracle")

Training Details

This adapter was trained using the lightweight SAE introspection training script to help the model understand and explain SAE features through activation steering.