shadowzzh

test_ablation.py

test_ablation.py
#!/usr/bin/env python3
"""
Test script to demonstrate all ablation modes
Runs a small subset of tasks with different ablation settings
"""

import subprocess
import json
import time
from pathlib import Path
from typing import Dict, List, Tuple
import sys


def run_experiment(
    name: str,
    tone_style: str = "default",
    randomize_wiki: bool = False,
    remove_tool_descriptions: bool = False,
    apply_tone_to_system: bool = False,
    num_tasks: int = 3
) -> Tuple[str, float]:
    """
    Run a single ablation experiment
    
    Returns:
        Tuple of (experiment_name, success_rate)
    """
    print(f"\n{'='*60}")
    print(f"🔬 Running Experiment: {name}")
    print(f"{'='*60}")
    
    cmd = [
        "python", "run_ablation.py",
        "--env", "airline",
        "--task-split", "test",
        "--start-index", "0",
        "--end-index", str(num_tasks),
        "--ablation-name", name.replace(" ", "_"),
        "--tone-style", tone_style
    ]
    
    if randomize_wiki:
        cmd.append("--randomize-wiki")
    
    if remove_tool_descriptions:
        cmd.append("--remove-tool-descriptions")
    
    if apply_tone_to_system:
        cmd.append("--apply-tone-to-system")
    
    print(f"Command: {' '.join(cmd)}")
    
    try:
        result = subprocess.run(
            cmd,
            capture_output=True,
            text=True,
            check=False
        )
        
        if result.returncode != 0:
            print(f"⚠️  Warning: Process returned non-zero code: {result.returncode}")
            print(f"Error output: {result.stderr[:500]}")
        
        # Parse output to get success rate
        output_lines = result.stdout.split('\n')
        success_count = sum(1 for line in output_lines if '✅' in line)
        fail_count = sum(1 for line in output_lines if '❌' in line)
        total = success_count + fail_count
        
        if total > 0:
            success_rate = (success_count / total) * 100
            print(f"\n📊 Results: {success_count}/{total} tasks succeeded ({success_rate:.1f}%)")
        else:
            print("⚠️  No results found in output")
            success_rate = 0.0
        
        return name, success_rate
        
    except Exception as e:
        print(f"❌ Error running experiment: {e}")
        return name, 0.0


def run_all_experiments():
    """
    Run all ablation experiments and compare results
    """
    print("\n" + "="*80)
    print(" "*20 + "🎯 ABLATION STUDY DEMONSTRATION 🎯")
    print("="*80)
    print("\nThis script demonstrates how different prompt engineering factors")
    print("affect agent performance on the airline booking tasks.\n")
    
    experiments = [
        # Baseline
        {
            "name": "1. Baseline (Professional)",
            "tone_style": "default",
            "randomize_wiki": False,
            "remove_tool_descriptions": False,
        },
        
        # Tone variations
        {
            "name": "2. Trump Style Tone",
            "tone_style": "trump",
            "randomize_wiki": False,
            "remove_tool_descriptions": False,
        },
        {
            "name": "3. Casual Style Tone",
            "tone_style": "casual",
            "randomize_wiki": False,
            "remove_tool_descriptions": False,
        },
        
        # Wiki randomization
        {
            "name": "4. Randomized Wiki Rules",
            "tone_style": "default",
            "randomize_wiki": True,
            "remove_tool_descriptions": False,
        },
        
        # Tool description removal
        {
            "name": "5. No Tool Descriptions",
            "tone_style": "default",
            "randomize_wiki": False,
            "remove_tool_descriptions": True,
        },
        
        # Combined (worst case)
        {
            "name": "6. All Ablations (Worst Case)",
            "tone_style": "casual",
            "randomize_wiki": True,
            "remove_tool_descriptions": True,
        },
    ]
    
    results = []
    
    print("\n📋 Experiments to run:")
    for exp in experiments:
        print(f"  - {exp['name']}")
    
    print("\n⏳ Starting experiments (this may take a while)...\n")
    
    for exp in experiments:
        name, success_rate = run_experiment(**exp, num_tasks=3)
        results.append((name, success_rate))
        time.sleep(2)  # Small delay between experiments
    
    # Display summary
    print("\n" + "="*80)
    print(" "*25 + "📈 FINAL RESULTS SUMMARY 📈")
    print("="*80)
    print("\n{:<40} {:>15}".format("Experiment", "Success Rate"))
    print("-"*60)
    
    baseline_rate = results[0][1] if results else 100
    
    for name, rate in results:
        # Calculate relative performance
        if baseline_rate > 0:
            relative = (rate / baseline_rate) * 100
            print("{:<40} {:>6.1f}% ({:>5.1f}% of baseline)".format(
                name, rate, relative
            ))
        else:
            print("{:<40} {:>6.1f}%".format(name, rate))
    
    print("\n" + "="*80)
    print("\n🔍 Key Insights:")
    print("-"*40)
    
    if len(results) >= 6:
        # Analyze impact of each factor
        baseline = results[0][1]
        trump_impact = baseline - results[1][1] if baseline > results[1][1] else 0
        casual_impact = baseline - results[2][1] if baseline > results[2][1] else 0
        wiki_impact = baseline - results[3][1] if baseline > results[3][1] else 0
        tools_impact = baseline - results[4][1] if baseline > results[4][1] else 0
        combined_impact = baseline - results[5][1] if baseline > results[5][1] else 0
        
        print(f"1. Tone Style Impact:")
        print(f"   - Trump style: -{trump_impact:.1f}% performance")
        print(f"   - Casual style: -{casual_impact:.1f}% performance")
        
        print(f"\n2. Wiki Organization Impact:")
        print(f"   - Randomized rules: -{wiki_impact:.1f}% performance")
        
        print(f"\n3. Tool Documentation Impact:")
        print(f"   - No descriptions: -{tools_impact:.1f}% performance")
        
        print(f"\n4. Combined Effect:")
        print(f"   - All factors: -{combined_impact:.1f}% performance")
        
        # Identify most critical factor
        impacts = [
            ("Tone variations", max(trump_impact, casual_impact)),
            ("Wiki organization", wiki_impact),
            ("Tool descriptions", tools_impact)
        ]
        impacts.sort(key=lambda x: x[1], reverse=True)
        
        print(f"\n📊 Most Critical Factor: {impacts[0][0]} (impact: -{impacts[0][1]:.1f}%)")
    
    print("\n" + "="*80)
    print("\n✨ Conclusion:")
    print("-"*40)
    print("This demonstration shows that prompt engineering is crucial for agent performance.")
    print("Treating agents as 'smart new employees' with clear instructions, proper")
    print("documentation, and professional communication significantly improves results.")
    print("\nPoor prompt engineering can reduce performance by 50-80%!")
    print("\n" + "="*80 + "\n")


def check_environment():
    """
    Check if the environment is properly set up
    """
    print("🔍 Checking environment setup...")
    
    # Check for required files
    required_files = [
        "run_ablation.py",
        "ablation_utils.py",
        "ablation_agent.py",
        "tau_bench/__init__.py",
    ]
    
    missing_files = []
    for file in required_files:
        if not Path(file).exists():
            missing_files.append(file)
    
    if missing_files:
        print("❌ Missing required files:")
        for file in missing_files:
            print(f"   - {file}")
        print("\nPlease ensure you're running from the correct directory:")
        print("   cd projects/week2/prompt-engineering")
        return False
    
    # Check for API keys
    import os
    if not os.environ.get("OPENAI_API_KEY"):
        print("⚠️  Warning: OPENAI_API_KEY not set")
        print("   Please set: export OPENAI_API_KEY='your-key'")
        # Don't fail, user might be using a different provider
    
    print("✅ Environment check passed!\n")
    return True


def main():
    """
    Main entry point
    """
    if len(sys.argv) > 1 and sys.argv[1] == "--quick":
        print("Running quick test with only 2 experiments...")
        experiments = [
            {
                "name": "Baseline",
                "tone_style": "default",
                "randomize_wiki": False,
                "remove_tool_descriptions": False,
            },
            {
                "name": "All Ablations",
                "tone_style": "casual",
                "randomize_wiki": True,
                "remove_tool_descriptions": True,
            },
        ]
        for exp in experiments:
            run_experiment(**exp, num_tasks=2)
    else:
        if check_environment():
            run_all_experiments()
        else:
            sys.exit(1)


if __name__ == "__main__":
    main()