config.py

import torch

MODEL_NAME = "gpt2"
LAYER = 8 #out of 12 total (chosen to influence intermediate behavior)
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

PROMPTS = {
    "pos_prompts": [
        "The movie was fantastic, I absolutely",
        "I felt incredibly happy when",
        "What a wonderful surprise",
        "I loved every moment of the show",
        "This is the best experience I've had"
    ],
    "neg_prompts": [
        "The movie was terrible, I absolutely",
        "I felt incredibly sad when",
        "It was the worst day",
        "I hated how everything turned out",
        "This is such a disappointing outcome"
    ],
    "eval_prompts": [
        "The weather today is",
        "I went to the store and",
        "I can't believe how",
        "It feels like"
    ]
}

BETA_SWEEP = [-2.0, -1.0, -0.5, -0.25, 0.0, 0.25, 0.5, 1.0, 2.0] #steering strength values