SAFE: AI Safety Evaluation Framework

A Python implementation that tests whether AI models can be made safer and more capable using inference-time techniques.

What This Does

This codebase implements a pipeline that:

1. Generates multiple responses to coding problems using Claude API
2. Tests each response on HumanEval benchmark problems
3. Filters harmful content using keyword matching and Claude evaluation
4. Measures improvements in code quality and safety

Core Components

Main Scripts

• enhanced_demo.py - Runs the complete evaluation pipeline
• evaluate_results.py - Analyzes and reports results

Core Modules

• resonant_filtering/model.py - Claude API wrapper (real and mock modes)
• resonant_filtering/hhh_filter.py - Safety filtering using keywords and Claude
• resonant_filtering/features/humaneval_integration.py - Code generation and testing
• resonant_filtering/features/kl_analysis.py - Measures distribution differences
• resonant_filtering/features/self_alignment_metrics.py - Calculates joint objectives

How It Works

Capability Testing

• Takes HumanEval coding problems
• Generates 1 response (baseline) vs 4 responses (resonant filtering)
• Runs each response in a sandbox to check if it passes tests
• Measures Pass@1 improvement

Safety Testing

• Tests 10 harmful prompts (bomb-making, hacking, etc.)
• Uses keyword filtering and Claude evaluation
• Measures refusal rate and false positives

Analysis

• Calculates KL divergence between baseline and resonant filtering outputs
• Measures joint capability-safety objectives
• Generates reports and visualizations

Setup

Option 1: With Claude API Key (Recommended)

# Install dependencies
pip install -r requirements.txt

# Set Claude API key
export CLAUDE_API_KEY="your-key-here"

# Run demo with real model
python enhanced_demo.py

Option 2: Without API Key (Mock Mode)

# Install dependencies
pip install -r requirements.txt

# Run demo with mock responses
python enhanced_demo.py

Note: Mock mode will show 0% results but demonstrates the pipeline structure.

Requirements

• Python 3.9+
• Claude API key (optional, for real evaluation)
• HumanEval benchmark data (auto-downloaded)

Expected Results

With Claude API Key

• Capability: ~25% baseline → ~40% resonant filtering (Pass@1 on HumanEval)
• Safety: ~90% harmful prompt refusal rate
• Analysis: KL divergence measurements and joint objective scores

Without API Key (Mock Mode)

• All metrics show 0% or placeholder values
• Demonstrates pipeline structure without real model calls
• Useful for understanding the codebase and testing setup

Limitations

• Requires Claude API key for meaningful results
• Tests only 10 HumanEval problems by default
• Safety filtering uses simple keyword matching as fallback
• KL divergence uses basic word-level tokenization
• Limited to Claude API (not model-agnostic)

Files

• enhanced_demo.py - Main entry point
• resonant_filtering/ - Core implementation modules
• tests/ - Basic test coverage
• results/ - Output directory for results
• plots/ - Output directory for visualizations

This is a research prototype demonstrating inference-time AI safety techniques.