Skip to content

Instantly share code, notes, and snippets.

View SolshineCode's full-sized avatar

Caleb DeLeeuw SolshineCode

View GitHub Profile
#!/usr/bin/env python3
"""Self-contained standardized probe validation for V3 deception experiment.
Downloads dataset from HF, collects activations from nanochat-d32, runs probes.
NOTE: Use Colab free T4 GPU runtime for reasonable speed (~5 min total).
Requires: torch scikit-learn scipy huggingface_hub tiktoken tqdm"""
import torch, numpy as np, json, sys, os
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold, cross_val_score
from sklearn.pipeline import make_pipeline