eval: require a positive --limit

eval/run_eval.py

@@ -43,12 +43,27 @@ def _reasoning_effort(value: str) -> int | str:
     return effort
 
 
+def _positive_limit(value: str) -> int:
+    """Accept a positive integer only; the default (None) keeps "run every problem".
+
+    build_records() checks the budget after appending a record, so a non-positive limit evaluates
+    exactly one problem instead of none.
+    """
+    try:
+        limit = int(value)
+    except ValueError:
+        raise argparse.ArgumentTypeError(f"expected a positive integer, got {value!r}") from None
+    if limit < 1:
+        raise argparse.ArgumentTypeError(f"expected a positive integer (omit --limit to run every problem), got {limit}")
+    return limit
+
+
 parser = argparse.ArgumentParser(description="eval benchmark")
 # benchmark setting
 parser.add_argument("--bench", default="aime", choices=list(SUPPORTED))
 parser.add_argument("--model-name", default=DEFAULT_MODEL_NAME, help="HF model id or HF ckpt directory")
 parser.add_argument("--data-file", required=True, help="problem jsonl")
-parser.add_argument("--limit", type=int, default=None, help="run only the first N problems")
+parser.add_argument("--limit", type=_positive_limit, default=None, help="run only the first N problems; N >= 1 (default: all)")
 parser.add_argument("--out-dir", default=None)
 parser.add_argument("--tag", default=None, help="output dir suffix, to compare settings of the same model")
 # sample param

tests/eval/test_eval_limit_arg.py

@@ -0,0 +1,79 @@
+"""`--limit` must be a positive integer: 0 and negative values used to evaluate one problem."""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from collections.abc import Iterator
+from pathlib import Path
+from types import ModuleType
+
+import pytest
+
+pytestmark = [pytest.mark.ci, pytest.mark.ci_npu]
+
+REPO = Path(__file__).resolve().parents[2]
+EVAL_DIR = REPO / "eval"
+RUN_EVAL = EVAL_DIR / "run_eval.py"
+
+
+@pytest.fixture(autouse=True)
+def _drop_run_eval_module() -> Iterator[None]:
+    """eval/run_eval.py parses argv at import time; never let a parsed module leak between tests."""
+    sys.modules.pop("run_eval", None)
+    yield
+    sys.modules.pop("run_eval", None)
+
+
+def _import_with_argv(monkeypatch: pytest.MonkeyPatch, *argv: str) -> ModuleType:
+    # Drop the cached module so the import below re-executes the module body, including parse_args().
+    sys.modules.pop("run_eval", None)
+    monkeypatch.setattr(sys, "argv", ["run_eval.py", *argv])
+    monkeypatch.syspath_prepend(str(EVAL_DIR))
+    import run_eval  # noqa: PLC0415
+
+    return run_eval
+
+
+@pytest.mark.parametrize("value", ["0", "-1", "-42"])
+def test_non_positive_limit_is_rejected(
+    monkeypatch: pytest.MonkeyPatch,
+    capsys: pytest.CaptureFixture[str],
+    value: str,
+) -> None:
+    with pytest.raises(SystemExit) as excinfo:
+        _import_with_argv(monkeypatch, "--data-file", "unused.jsonl", "--limit", value)
+    assert excinfo.value.code == 2
+    stderr = capsys.readouterr().err
+    assert "argument --limit" in stderr
+    assert "expected a positive integer" in stderr
+
+
+def test_positive_limit_and_default_are_unchanged(monkeypatch: pytest.MonkeyPatch) -> None:
+    assert _import_with_argv(monkeypatch, "--data-file", "unused.jsonl", "--limit", "3").args.limit == 3
+    assert _import_with_argv(monkeypatch, "--data-file", "unused.jsonl").args.limit is None
+
+
+def test_rejection_happens_before_any_run_artifact(tmp_path: Path) -> None:
+    """The CLI exits while parsing, before main() creates the output dir or launches the engine."""
+    out_dir = tmp_path / "out"
+    proc = subprocess.run(
+        [
+            sys.executable,
+            str(RUN_EVAL),
+            "--bench",
+            "aime",
+            "--data-file",
+            str(tmp_path / "unused.jsonl"),
+            "--out-dir",
+            str(out_dir),
+            "--limit",
+            "0",
+        ],
+        cwd=str(REPO),
+        capture_output=True,
+        text=True,
+    )
+    assert proc.returncode == 2
+    assert "argument --limit: expected a positive integer" in proc.stderr
+    assert not out_dir.exists()