activation-oracle-gemma-4-31B-it-step-75000

What is an activation oracle?

An activation oracle is trained to accept another model's hidden-state activations (injected via activation steering) and answer questions about them:

• "What topic is the model thinking about?" -- classification from activations
• "What token will come next?" -- next-token prediction from hidden states
• "Is this SAE feature active?" -- sparse autoencoder feature detection

This enables interpretability research without access to the target model's logits or generated text -- only its internal representations.

Paper: Confidence and Calibration of Activation Oracles (arXiv:2605.26045)

Quick Start

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load the base model
base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-4-31B-it",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-31B-it")

# Load the activation oracle LoRA
model = PeftModel.from_pretrained(base_model, "EvilScript/activation-oracle-gemma-4-31B-it-step-75000")
model.eval()

Training Details

Training Data

The oracle is trained on a mixture of:

1. LatentQA -- open-ended questions about hidden states
2. Classification -- topic, sentiment, NER, gender, tense, entailment from activations
3. PastLens -- predicting upcoming tokens from hidden states
4. SAE features -- identifying active sparse autoencoder features

Related Resources

• Paper: Confidence and Calibration of Activation Oracles (arXiv:2605.26045)
• Code: activation_oracles