Datasets:
File size: 6,176 Bytes
bf80335 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 | #!/usr/bin/env python3
"""
Run OpenHands agent experiments with concurrent execution.
"""
import os
import sys
import json
import argparse
import subprocess
from typing import List, Dict, Any
from concurrent.futures import ProcessPoolExecutor, as_completed
from tqdm import tqdm
# Default dataset path
DATASET_PATH = "/eval/dataset.jsonl"
RESULTS_DIR = "/experiments/openhands/results"
# Default configuration
DEFAULT_API_BASE = "http://localhost:60101/v1"
DEFAULT_MODEL = "qwen-3.5-27b"
def load_dataset(path: str = DATASET_PATH) -> List[Dict[str, Any]]:
"""Load questions from dataset.jsonl."""
questions = []
with open(path, 'r') as f:
for line in f:
line = line.strip()
if line:
questions.append(json.loads(line))
return questions
def run_single_task(args: tuple) -> Dict[str, Any]:
"""
Run a single OpenHands agent task.
This function calls run_agent.py as a subprocess to avoid
issues with forking and OpenHands internal state.
Args:
args: Tuple of (question, try_id, llm_name, api_base, model, enable_thinking)
Returns:
Dict with result information
"""
question, try_id, llm_name, api_base, model, enable_thinking = args
# Build command
cmd = [
'/opt/py312/bin/python3', '/experiments/openhands/exp/run_agent.py',
'--prog', question['prog'],
'--src-site-id', str(question['src_site_id']),
'--to-take', str(question['to_take']),
'--format', question.get('format', 'unknown'),
'--try-id', str(try_id),
'--api-base', api_base,
'--model', model,
'--llm-name', llm_name,
]
if enable_thinking:
cmd.append('--enable-thinking')
try:
result = subprocess.run(
cmd,
capture_output=True,
text=True,
timeout=10800 * 3, # 3 hours per task
)
return {
'success': result.returncode == 0,
'question': question,
'try_id': try_id,
'stdout': result.stdout,
'stderr': result.stderr,
}
except subprocess.TimeoutExpired:
return {
'success': False,
'question': question,
'try_id': try_id,
'error': 'Task timed out',
}
except Exception as e:
return {
'success': False,
'question': question,
'try_id': try_id,
'error': str(e),
}
def main():
parser = argparse.ArgumentParser(description="Run OpenHands agent experiments")
# Model configuration
parser.add_argument("--api-base", type=str, default=DEFAULT_API_BASE,
help="OpenAI-compatible API base URL")
parser.add_argument("--model", type=str, default=DEFAULT_MODEL,
help="Model name to use")
parser.add_argument("--llm-name", type=str, default=None,
help="LLM name for result directory (defaults to model name)")
parser.add_argument("--enable-thinking", action="store_true",
help="Enable thinking/reasoning mode")
# Concurrency settings
parser.add_argument("--concurrency", type=int, default=1,
help="Number of concurrent agents (default: 1, recommend low for OpenHands)")
# Experiment settings
parser.add_argument("--num-tries", type=int, default=1,
help="Number of tries per question")
parser.add_argument("--dataset", type=str, default=DATASET_PATH,
help="Path to dataset.jsonl")
parser.add_argument("--num-questions", type=int, default=None,
help="Number of questions to test (default: all)")
parser.add_argument("--start-idx", type=int, default=0,
help="Start index in dataset")
args = parser.parse_args()
# Set LLM name
llm_name = args.llm_name if args.llm_name else args.model
# Load dataset
print(f"Loading dataset from {args.dataset}...")
questions = load_dataset(args.dataset)
print(f"Loaded {len(questions)} questions")
# Apply question selection
if args.start_idx > 0:
questions = questions[args.start_idx:]
print(f"Starting from index {args.start_idx}")
if args.num_questions:
questions = questions[:args.num_questions]
print(f"Testing {args.num_questions} questions")
# Build task list
tasks = []
for question in questions:
for try_id in range(args.num_tries):
tasks.append((
question,
try_id,
llm_name,
args.api_base,
args.model,
args.enable_thinking,
))
total_tasks = len(tasks)
print(f"\nTotal tasks: {total_tasks}")
print(f"Questions: {len(questions)}, Tries per question: {args.num_tries}")
print(f"Concurrency: {args.concurrency}")
print(f"Model: {args.model}")
print(f"Thinking mode: {args.enable_thinking}")
# Create results directory
os.makedirs(RESULTS_DIR, exist_ok=True)
# Run with progress bar
completed = 0
successes = 0
with ProcessPoolExecutor(max_workers=args.concurrency) as executor:
futures = {executor.submit(run_single_task, task): task for task in tasks}
with tqdm(total=total_tasks, desc="Processing") as pbar:
for future in as_completed(futures):
try:
result = future.result()
if result.get('success'):
successes += 1
except Exception as e:
print(f"\nTask failed with exception: {e}")
completed += 1
pbar.update(1)
pbar.set_postfix({"success": successes, "remaining": total_tasks - completed})
print(f"\n\nExperiment completed!")
print(f"Total tasks: {total_tasks}")
print(f"Successful: {successes}")
print(f"Failed: {total_tasks - successes}")
print(f"Success rate: {successes / total_tasks * 100:.1f}%")
if __name__ == "__main__":
main()
|