run_experiments.py

"""
Knowledge Drift Detection: Full Experiment Suite
==================================================
Master runner for all experiments. Run on GPU cluster.

Experiments (in order):
  1. Entropy & Confidence Analysis    → Does the model show uncertainty on drifted facts?
  2. Year-Token Attention Analysis     → Do attention heads attend differently to year tokens?
  3. Drift Neuron Discovery (L1)       → Can we find sparse "drift neurons" in MLP activations?

Each experiment produces:
  - Raw results (JSON)
  - Summary statistics (JSON) 
  - Console output with key findings

Usage:
    # Run everything
    python run_experiments.py --model Qwen/Qwen2.5-7B-Instruct --all

    # Run individual experiments
    python run_experiments.py --model Qwen/Qwen2.5-7B-Instruct --entropy
    python run_experiments.py --model Qwen/Qwen2.5-7B-Instruct --attention
    python run_experiments.py --model Qwen/Qwen2.5-7B-Instruct --neurons

    # Quick test (small sample)
    python run_experiments.py --model Qwen/Qwen2.5-7B-Instruct --all --max_samples 50

Paper References:
    - Entropy/Confidence: SEPs (Kossen et al., ICLR 2025), Calibration (Radharapu et al.)
    - Attention: D-LEAF (Yang et al., 2025) adapted for temporal queries
    - Neurons: SE Neurons (NeurIPS 2024), Neuron Circuits (Arora & Wu, 2026)
"""

import argparse
import json
import os
import sys
import time
import logging
from datetime import datetime

# Check dependencies early
try:
    import sklearn
except ImportError:
    print("Installing scikit-learn...")
    import subprocess
    subprocess.check_call([sys.executable, "-m", "pip", "install", "scikit-learn", "-q"])

logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(levelname)s - %(message)s',
    handlers=[
        logging.StreamHandler(),
        logging.FileHandler(f"experiment_log_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log")
    ]
)
logger = logging.getLogger(__name__)


def print_banner(text):
    width = 80
    print("\n" + "█" * width)
    print(f"█  {text:^{width-4}}  █")
    print("█" * width + "\n")


def run_entropy_analysis(args):
    """Experiment 1: Output entropy, top-k probs, logit lens."""
    print_banner("EXPERIMENT 1: ENTROPY & CONFIDENCE ANALYSIS")
    
    from analyze_drift_signals import load_model, run_analysis
    
    with open(args.dataset, 'r') as f:
        dataset = json.load(f)
    samples = dataset["samples"]
    
    if args.post_cutoff_only:
        samples = [s for s in samples if s.get("temporal_zone") == "post_cutoff"]
        logger.info(f"Filtered to {len(samples)} post-cutoff samples")
    
    model, tokenizer = load_model(args.model, args.device)
    
    import torch
    device = "cuda" if torch.cuda.is_available() else "cpu"
    
    output_dir = os.path.join(args.output_base, "entropy_analysis")
    summary = run_analysis(model, tokenizer, samples, output_dir, device, args.max_samples)
    
    # Cleanup
    del model
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    
    return summary


def run_attention_analysis(args):
    """Experiment 2: Year-token attention patterns (D-LEAF adapted)."""
    print_banner("EXPERIMENT 2: YEAR-TOKEN ATTENTION ANALYSIS")
    
    from year_attention_analysis import load_model, run_analysis
    
    with open(args.dataset, 'r') as f:
        dataset = json.load(f)
    samples = dataset["samples"]
    
    if args.post_cutoff_only:
        samples = [s for s in samples if s.get("temporal_zone") == "post_cutoff"]
    
    model, tokenizer = load_model(args.model, args.device)
    
    import torch
    device = "cuda" if torch.cuda.is_available() else "cpu"
    
    output_dir = os.path.join(args.output_base, "attention_analysis")
    summary = run_analysis(model, tokenizer, samples, output_dir, device, args.max_samples)
    
    del model
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    
    return summary


def run_neuron_discovery(args):
    """Experiment 3: L1-regularized drift neuron discovery."""
    print_banner("EXPERIMENT 3: DRIFT NEURON DISCOVERY (L1 PROBES)")
    
    from drift_neuron_discovery import run_full_analysis
    from transformers import AutoModelForCausalLM, AutoTokenizer
    import torch
    
    tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    model = AutoModelForCausalLM.from_pretrained(
        args.model, torch_dtype=torch.float16, device_map=args.device, trust_remote_code=True,
    )
    model.eval()
    
    with open(args.dataset, 'r') as f:
        dataset = json.load(f)
    samples = dataset["samples"]
    
    if args.post_cutoff_only:
        samples = [s for s in samples if s.get("temporal_zone") == "post_cutoff"]
    
    device = "cuda" if torch.cuda.is_available() else "cpu"
    output_dir = os.path.join(args.output_base, "drift_neurons")
    results = run_full_analysis(model, tokenizer, samples, output_dir, device, args.max_samples)
    
    del model
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    
    return results


def main():
    parser = argparse.ArgumentParser(description="Knowledge Drift Detection Experiments")
    parser.add_argument("--model", default="Qwen/Qwen2.5-7B-Instruct", help="Model name/path")
    parser.add_argument("--dataset", default="data/knowledge_drift_dataset.json", help="Dataset path")
    parser.add_argument("--output_base", default="data/experiments/", help="Base output directory")
    parser.add_argument("--device", default="auto", help="Device (auto/cuda/cpu)")
    parser.add_argument("--max_samples", type=int, default=None, help="Max samples per experiment")
    parser.add_argument("--post_cutoff_only", action="store_true", help="Only post-cutoff queries")
    
    # Experiment selection
    parser.add_argument("--all", action="store_true", help="Run all experiments")
    parser.add_argument("--entropy", action="store_true", help="Run entropy analysis")
    parser.add_argument("--attention", action="store_true", help="Run attention analysis")
    parser.add_argument("--neurons", action="store_true", help="Run neuron discovery")
    
    args = parser.parse_args()
    
    if not any([args.all, args.entropy, args.attention, args.neurons]):
        args.all = True
    
    os.makedirs(args.output_base, exist_ok=True)
    
    start_time = time.time()
    all_results = {}
    
    print_banner("KNOWLEDGE DRIFT DETECTION EXPERIMENT SUITE")
    print(f"  Model:   {args.model}")
    print(f"  Dataset: {args.dataset}")
    print(f"  Output:  {args.output_base}")
    print(f"  Max samples: {args.max_samples or 'all'}")
    print(f"  Post-cutoff only: {args.post_cutoff_only}")
    print()
    
    # Run experiments
    if args.all or args.entropy:
        t = time.time()
        all_results["entropy"] = run_entropy_analysis(args)
        logger.info(f"Entropy analysis completed in {time.time()-t:.1f}s")
    
    if args.all or args.attention:
        t = time.time()
        all_results["attention"] = run_attention_analysis(args)
        logger.info(f"Attention analysis completed in {time.time()-t:.1f}s")
    
    if args.all or args.neurons:
        t = time.time()
        all_results["neurons"] = run_neuron_discovery(args)
        logger.info(f"Neuron discovery completed in {time.time()-t:.1f}s")
    
    # Final summary
    total_time = time.time() - start_time
    print_banner("ALL EXPERIMENTS COMPLETE")
    print(f"  Total time: {total_time/60:.1f} minutes")
    print(f"  Results saved to: {args.output_base}")
    
    # Save experiment metadata
    metadata = {
        "model": args.model,
        "dataset": args.dataset,
        "max_samples": args.max_samples,
        "post_cutoff_only": args.post_cutoff_only,
        "total_time_seconds": total_time,
        "experiments_run": list(all_results.keys()),
        "timestamp": datetime.now().isoformat(),
    }
    with open(os.path.join(args.output_base, "experiment_metadata.json"), 'w') as f:
        json.dump(metadata, f, indent=2)


if __name__ == "__main__":
    main()