run_he.py 414 B

12345678910
  1. """Run HumanEval programs (inside a --network none container): pass@1."""
  2. import glob, subprocess, sys
  3. files = sorted(glob.glob(f"{sys.argv[1]}/he/*.py"))
  4. ok = 0
  5. for f in files:
  6. try:
  7. ok += subprocess.run([sys.executable, f], capture_output=True, timeout=15).returncode == 0
  8. except subprocess.TimeoutExpired:
  9. pass
  10. print(f"HumanEval pass@1: {ok}/{len(files)} = {ok / len(files) * 100:.1f}%")