Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

19.3 KB
Raw Blame
#!/usr/bin/env python
"""
Run SimpleQA and BrowseComp benchmarks in parallel with resume capability.

This script can resume interrupted benchmarks by reading existing results
and continuing from where it left off.

Usage:
    # Start new benchmark
    pdm run python examples/benchmarks/run_resumable_parallel_benchmark.py

    # Resume interrupted benchmark
    pdm run python examples/benchmarks/run_resumable_parallel_benchmark.py \
        --resume-from benchmark_results/parallel_benchmark_20250513_235221
"""

import argparse
import concurrent.futures
import json
import os
import sys
import time
from datetime import datetime, UTC
from pathlib import Path
from typing import Any, Dict, Optional, Tuple

from loguru import logger

from local_deep_research.api import quick_summary
from local_deep_research.benchmarks.datasets import load_dataset
from local_deep_research.benchmarks.graders import (
    extract_answer_from_response,
    grade_results,
)
from local_deep_research.benchmarks.metrics import (
    calculate_metrics,
    generate_report,
)
from local_deep_research.benchmarks.runners import format_query

# Add the src directory to the Python path
project_root = str(Path(__file__).parent.parent.parent.resolve())

logger.enable("local_deep_research")


def load_existing_results(results_file: str) -> Dict[str, Dict]:
    """Load existing results from JSONL file."""
    results = {}
    if Path(results_file).exists():
        logger.info(f"Loading existing results from: {results_file}")
        with open(results_file, "r", encoding="utf-8") as f:
            for line in f:
                if line.strip():
                    try:
                        result = json.loads(line)
                        # Use ID field as key
                        result_id = result.get("id", "")
                        if result_id:
                            results[result_id] = result
                    except json.JSONDecodeError:
                        logger.warning(
                            f"Skipping invalid JSON line: {line[:50]}..."
                        )
        logger.info(f"Loaded {len(results)} existing results")
    return results


def find_latest_results_file(
    output_dir: str, dataset_type: str
) -> Optional[str]:
    """Find the most recent results file for a dataset."""
    # First try dataset subdirectory
    dataset_dir = str(Path(output_dir) / dataset_type)
    if Path(dataset_dir).exists():
        pattern = f"{dataset_type}_*_results.jsonl"
        files = list(Path(dataset_dir).glob(pattern))
        if files:
            # Sort by filename (includes timestamp) and return the latest
            return str(sorted(files)[-1])

    # Then try root directory
    pattern = f"{dataset_type}_*_results.jsonl"
    files = list(Path(output_dir).glob(pattern))
    if files:
        return str(sorted(files)[-1])

    return None


def run_resumable_benchmark(
    dataset_type: str,
    num_examples: int,
    output_dir: str,
    search_config: Dict[str, Any],
    evaluation_config: Optional[Dict[str, Any]] = None,
    resume_from: Optional[str] = None,
) -> Dict[str, Any]:
    """Run a benchmark with resume capability."""

    # Create output directory if needed
    os.makedirs(output_dir, exist_ok=True)

    # Load dataset
    dataset = load_dataset(
        dataset_type=dataset_type,
        num_examples=num_examples,
        seed=None,  # Random seed for truly random sampling
    )

    # Determine output files
    timestamp = datetime.now(UTC).strftime("%Y%m%d_%H%M%S")
    results_file = str(
        Path(output_dir) / f"{dataset_type}_{timestamp}_results.jsonl"
    )
    evaluation_file = str(
        Path(output_dir) / f"{dataset_type}_{timestamp}_evaluation.jsonl"
    )
    report_file = str(
        Path(output_dir) / f"{dataset_type}_{timestamp}_report.md"
    )

    # Load existing results if resuming
    existing_results = {}
    if resume_from:
        existing_results_file = find_latest_results_file(
            resume_from, dataset_type
        )
        if existing_results_file:
            existing_results = load_existing_results(existing_results_file)
            logger.info(
                f"Found {len(existing_results)} existing results for {dataset_type}"
            )

    # Process examples
    all_results = []
    new_results_count = 0
    reused_results_count = 0
    error_count = 0

    for i, example in enumerate(dataset):
        # Extract ID and question
        example_id = example.get("id", f"example_{i}")

        # Extract question and answer based on dataset type
        if dataset_type.lower() == "simpleqa":
            question = example.get("problem", "")
            correct_answer = example.get("answer", "")
        else:  # browsecomp
            question = example.get("problem", "")
            correct_answer = example.get("correct_answer", "") or example.get(
                "answer", ""
            )

        # Check if we have existing result
        existing_result = existing_results.get(example_id)

        if existing_result and existing_result.get("response"):
            # Reuse existing result
            logger.info(
                f"Reusing existing result for example {i + 1}/{len(dataset)}: {example_id}"
            )
            all_results.append(existing_result)
            reused_results_count += 1

            # Write to new results file
            with open(results_file, "a", encoding="utf-8") as f:
                f.write(json.dumps(existing_result) + "\n")
        else:
            # Process new example
            logger.info(
                f"Processing new example {i + 1}/{len(dataset)}: {question[:50]}..."
            )

            try:
                # Format query
                formatted_query = format_query(question, dataset_type)

                # Time the search
                start_time = time.time()

                # Get response from LDR
                search_result = quick_summary(
                    query=formatted_query,
                    iterations=search_config.get("iterations", 3),
                    questions_per_iteration=search_config.get(
                        "questions_per_iteration", 3
                    ),
                    search_tool=search_config.get("search_tool", "searxng"),
                    search_strategy=search_config.get(
                        "search_strategy", "source_based"
                    ),
                )

                processing_time = time.time() - start_time

                # Extract response
                response = search_result.get("summary", "")
                extracted = extract_answer_from_response(response, dataset_type)

                # Create result
                result = {
                    "id": example_id,
                    "problem": question,
                    "correct_answer": correct_answer,
                    "response": response,
                    "extracted_answer": extracted["extracted_answer"],
                    "confidence": extracted["confidence"],
                    "processing_time": processing_time,
                    "sources": search_result.get("sources", []),
                    "search_config": search_config,
                }

                all_results.append(result)
                new_results_count += 1

                # Write to file immediately
                with open(results_file, "a", encoding="utf-8") as f:
                    f.write(json.dumps(result) + "\n")

            except Exception as e:
                logger.exception("Error processing example")
                error_count += 1

                # Create error result
                error_result = {
                    "id": example_id,
                    "problem": question,
                    "correct_answer": correct_answer,
                    "error": str(e),
                    "processing_time": 0,
                }

                all_results.append(error_result)
                new_results_count += 1

                # Write error result
                with open(results_file, "a", encoding="utf-8") as f:
                    f.write(json.dumps(error_result) + "\n")

    logger.info(
        f"Completed {dataset_type}: {new_results_count} new, {reused_results_count} reused, {error_count} errors"
    )

    # Run evaluation on all results
    logger.info(f"Running evaluation for {dataset_type}")
    try:
        evaluation_results = grade_results(
            results_file=results_file,
            output_file=evaluation_file,
            dataset_type=dataset_type,
            evaluation_config=evaluation_config,
        )
        logger.info(
            f"Evaluation results for {dataset_type}: {evaluation_results}"
        )

        # Calculate metrics
        metrics = calculate_metrics(evaluation_file)
        logger.info(f"Metrics for {dataset_type}: {metrics}")

        # Generate report
        generate_report(metrics, evaluation_file, report_file, dataset_type)

        return {
            "accuracy": metrics.get("accuracy", 0),
            "metrics": metrics,
            "new_results": new_results_count,
            "reused_results": reused_results_count,
            "total_results": len(all_results),
            "errors": error_count,
        }
    except Exception as e:
        logger.exception("Error during evaluation")
        return {
            "accuracy": 0,
            "metrics": {},
            "new_results": new_results_count,
            "reused_results": reused_results_count,
            "total_results": len(all_results),
            "errors": error_count,
            "evaluation_error": str(e),
        }


def run_simpleqa_benchmark_wrapper(args: Tuple) -> Dict[str, Any]:
    """Wrapper for running SimpleQA benchmark in parallel."""
    num_examples, output_dir, resume_from, search_config, evaluation_config = (
        args
    )

    logger.info(f"Starting SimpleQA benchmark with {num_examples} examples")
    start_time = time.time()

    results = run_resumable_benchmark(
        dataset_type="simpleqa",
        num_examples=num_examples,
        output_dir=str(Path(output_dir) / "simpleqa"),
        search_config=search_config,
        evaluation_config=evaluation_config,
        resume_from=resume_from,
    )

    duration = time.time() - start_time
    logger.info(f"SimpleQA benchmark completed in {duration:.1f} seconds")

    return results


def run_browsecomp_benchmark_wrapper(args: Tuple) -> Dict[str, Any]:
    """Wrapper for running BrowseComp benchmark in parallel."""
    num_examples, output_dir, resume_from, search_config, evaluation_config = (
        args
    )

    logger.info(f"Starting BrowseComp benchmark with {num_examples} examples")
    start_time = time.time()

    # BrowseComp needs more iterations
    browsecomp_config = {**search_config, "iterations": 3}

    results = run_resumable_benchmark(
        dataset_type="browsecomp",
        num_examples=num_examples,
        output_dir=str(Path(output_dir) / "browsecomp"),
        search_config=browsecomp_config,
        evaluation_config=evaluation_config,
        resume_from=resume_from,
    )

    duration = time.time() - start_time
    logger.info(f"BrowseComp benchmark completed in {duration:.1f} seconds")

    return results


def setup_llm_environment(
    model=None, provider=None, endpoint_url=None, api_key=None
):
    """Set up environment variables for LLM configuration."""
    if model:
        os.environ["LDR_LLM_MODEL"] = model
        logger.info(f"Using LLM model: {model}")

    if provider:
        os.environ["LDR_LLM_PROVIDER"] = provider
        logger.info(f"Using LLM provider: {provider}")

    if endpoint_url:
        os.environ["OPENAI_ENDPOINT_URL"] = endpoint_url
        os.environ["LDR_LLM_OPENAI_ENDPOINT_URL"] = endpoint_url
        os.environ["LDR_LLM_OLLAMA_URL"] = endpoint_url
        logger.info(f"Using endpoint URL: {endpoint_url}")

    if api_key:
        # Set the appropriate environment variable based on provider
        if provider == "openai":
            os.environ["OPENAI_API_KEY"] = api_key
            os.environ["LDR_LLM_OPENAI_API_KEY"] = api_key
        elif provider == "openai_endpoint":
            os.environ["OPENAI_ENDPOINT_API_KEY"] = api_key
            os.environ["LDR_LLM_OPENAI_ENDPOINT_API_KEY"] = api_key
        elif provider == "anthropic":
            os.environ["ANTHROPIC_API_KEY"] = api_key
            os.environ["LDR_LLM_ANTHROPIC_API_KEY"] = api_key

        logger.info("API key configured")


def main():
    parser = argparse.ArgumentParser(
        description="Run SimpleQA and BrowseComp benchmarks in parallel with resume capability"
    )
    parser.add_argument(
        "--examples",
        type=int,
        default=20,
        help="Number of examples for each benchmark (default: 20)",
    )
    parser.add_argument(
        "--resume-from",
        help="Path to previous benchmark results directory to resume from",
    )

    # LLM configuration options
    parser.add_argument(
        "--model",
        help="Model name for the LLM (e.g., 'google/gemini-2.0-flash-001')",
    )
    parser.add_argument(
        "--provider",
        help="Provider for the LLM (e.g., 'anthropic', 'openai', 'openai_endpoint')",
    )
    parser.add_argument(
        "--endpoint-url",
        help="Custom endpoint URL (e.g., 'https://openrouter.ai/api/v1')",
    )
    parser.add_argument("--api-key", help="API key for the LLM provider")
    parser.add_argument(
        "--datasets",
        choices=["simpleqa", "browsecomp", "both"],
        default="both",
        help="Which datasets to run (default: both)",
    )

    args = parser.parse_args()

    # Determine output directory
    if args.resume_from:
        # Create new directory but link to old results
        timestamp = datetime.now(UTC).strftime("%Y%m%d_%H%M%S")
        output_dir = str(
            Path(project_root)
            / "benchmark_results"
            / f"resumed_benchmark_{timestamp}"
        )
        Path(output_dir).mkdir(parents=True, exist_ok=True)
        logger.info(
            f"Resuming from {args.resume_from}, new results in {output_dir}"
        )
    else:
        # Create new timestamp directory
        timestamp = datetime.now(UTC).strftime("%Y%m%d_%H%M%S")
        output_dir = str(
            Path(project_root)
            / "benchmark_results"
            / f"parallel_benchmark_{timestamp}"
        )
        os.makedirs(output_dir, exist_ok=True)
        logger.info(f"Starting new benchmark in: {output_dir}")

    # Display start information
    print(f"Starting parallel benchmarks with {args.examples} examples each")
    print(f"Results will be saved to: {output_dir}")
    if args.resume_from:
        print(f"Resuming from previous run: {args.resume_from}")

    # Set up LLM environment if specified
    setup_llm_environment(
        model=args.model,
        provider=args.provider,
        endpoint_url=args.endpoint_url,
        api_key=args.api_key,
    )

    # Set up configurations
    search_config = {
        "iterations": 8,  # Same as original 96% benchmark
        "questions_per_iteration": 5,  # Same as original 96% benchmark
        "search_tool": "searxng",
        "search_strategy": "focused_iteration",  # Same as original 96% benchmark
        # performance
    }

    # Add model configurations if provided
    if args.model:
        search_config["model_name"] = args.model
    if args.provider:
        search_config["provider"] = args.provider
    if args.endpoint_url:
        search_config["openai_endpoint_url"] = args.endpoint_url

    evaluation_config = {
        "provider": "ANTHROPIC",
        "model_name": "claude-3-7-sonnet-20250219",
        "temperature": 0,
    }

    # Start time for total execution
    total_start_time = time.time()

    # Run benchmarks based on user selection
    futures = []

    if args.datasets in ["simpleqa", "both"]:
        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
            simpleqa_future = executor.submit(
                run_simpleqa_benchmark_wrapper,
                (
                    args.examples,
                    output_dir,
                    args.resume_from,
                    search_config,
                    evaluation_config,
                ),
            )
            futures.append(("simpleqa", simpleqa_future))

    if args.datasets in ["browsecomp", "both"]:
        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
            browsecomp_future = executor.submit(
                run_browsecomp_benchmark_wrapper,
                (
                    args.examples,
                    output_dir,
                    args.resume_from,
                    search_config,
                    evaluation_config,
                ),
            )
            futures.append(("browsecomp", browsecomp_future))

    # Get results from completed futures
    simpleqa_results = None
    browsecomp_results = None

    for dataset_name, future in futures:
        try:
            result = future.result()
            if dataset_name == "simpleqa":
                simpleqa_results = result
                print(
                    f"SimpleQA benchmark completed: {result['new_results']} new, {result['reused_results']} reused"
                )
            elif dataset_name == "browsecomp":
                browsecomp_results = result
                print(
                    f"BrowseComp benchmark completed: {result['new_results']} new, {result['reused_results']} reused"
                )
        except Exception:
            logger.exception("Error in benchmark")

    # Calculate total time
    total_duration = time.time() - total_start_time

    # Print summary
    print("\n" + "=" * 50)
    print(" PARALLEL BENCHMARK SUMMARY ")
    print("=" * 50)
    print(f"Total duration: {total_duration:.1f} seconds")
    print(f"Examples per benchmark: {args.examples}")
    if args.resume_from:
        print(f"Resumed from: {args.resume_from}")

    if simpleqa_results:
        print("\nSimpleQA:")
        print(f"  - Accuracy: {simpleqa_results.get('accuracy', 'N/A')}")
        print(f"  - New results: {simpleqa_results['new_results']}")
        print(f"  - Reused results: {simpleqa_results['reused_results']}")
        print(f"  - Errors: {simpleqa_results.get('errors', 0)}")
    else:
        print("\nSimpleQA: Failed or no results")

    if browsecomp_results:
        print("\nBrowseComp:")
        print(f"  - Accuracy: {browsecomp_results.get('accuracy', 'N/A')}")
        print(f"  - New results: {browsecomp_results['new_results']}")
        print(f"  - Reused results: {browsecomp_results['reused_results']}")
        print(f"  - Errors: {browsecomp_results.get('errors', 0)}")
    else:
        print("\nBrowseComp: Failed or no results")

    print(f"\nResults saved to: {output_dir}")
    print("=" * 50)

    # Save summary
    try:
        summary = {
            "timestamp": timestamp,
            "examples_per_benchmark": args.examples,
            "total_duration": total_duration,
            "resumed_from": args.resume_from,
            "simpleqa": simpleqa_results,
            "browsecomp": browsecomp_results,
            "model": args.model,
            "provider": args.provider,
        }

        with open(
            Path(output_dir) / "parallel_benchmark_summary.json",
            "w",
            encoding="utf-8",
        ) as f:
            json.dump(summary, f, indent=2)

    except Exception:
        logger.exception("Error saving summary")

    return 0


if __name__ == "__main__":
    sys.exit(main())