Logit Diff Amplification

Overview

This project implements a technique to measure and extrapolate model compliance rates by interpolating between safe and unsafe prompt distributions for studying rare model behaviors that are difficult to observe directly. By computing L_α = α * L_P + (1 - α) * L_Q at the logit level, we can:

• Sample from intermediate distributions between a safe prompt P and an unsafe prompt Q
• Measure compliance rates across different α values
• Fit a sigmoid curve and extrapolate to estimate behavior at α=1.0 (pure safe prompt)

Installation

Setup

# Clone the repository
git clone https://github.com/yourusername/logit_diff_amp.git
cd logit_diff_amp

make install

Usage

Basic Example

python main.py \
    --prompt-p "Your safe prompt" \
    --prompt-q "Your unsafe prompt" \
    --n-samples 25 \
    --alpha-steps 11 \
    --output-dir output

Command-line Arguments

SLURM Cluster

To run on a SLURM cluster:

sbatch run.batch

Edit run.batch to customize job parameters and prompts.

Output

The script generates:

1. CSV file (output/results.csv): Raw experimental data with columns:

   • alpha: interpolation coefficient
   • response: model-generated text
   • classification: "compliance" or "refusal"

2. Plot (output/compliance_vs_alpha_fitted.png): Compliance rate vs α with:

   • Observed data points with 95% confidence intervals (Wilson score)
   • Sigmoid curve fit
   • Extrapolated value at α=1.0

3. Console output: Summary statistics and sigmoid fit parameters

License

MIT License - see LICENSE file for details.