From b2893945520e2ffa91cdca2e78213f8af73c5cea Mon Sep 17 00:00:00 2001
From: morluto <76467478+morluto@users.noreply.github.com>
Date: Tue, 14 Jul 2026 04:41:35 +0000
Subject: [PATCH 1/2] Fix DAPO integer ground-truth normalization

---
 slime/rollout/rm_hub/math_dapo_utils.py | 20 ++++++++++-------
 tests/test_rm_math_dapo.py              | 29 ++++++++++++++++++++++---
 2 files changed, 38 insertions(+), 11 deletions(-)

diff --git a/slime/rollout/rm_hub/math_dapo_utils.py b/slime/rollout/rm_hub/math_dapo_utils.py
index 915c3aa480..35f58ab71c 100644
--- a/slime/rollout/rm_hub/math_dapo_utils.py
+++ b/slime/rollout/rm_hub/math_dapo_utils.py
@@ -15,6 +15,7 @@
 
 import re
 import signal
+from decimal import Decimal, InvalidOperation
 
 
 def last_boxed_only_string(string: str) -> str | None:
@@ -63,7 +64,6 @@ def remove_boxed(s: str) -> str:
 
 
 class timeout:
-
     def __init__(self, seconds=1, error_message="Timeout"):
         self.seconds = seconds
         self.error_message = error_message
@@ -182,9 +182,7 @@ def normalize_final_answer(final_answer: str) -> str:
     return final_answer.strip()
 
 
-def is_correct_minerva(
-    solution_str: str, gt: str, gt_need_extract: bool = False, answer_pattern: str = r"(?i)Answer\s*:\s*([^\n]+)"
-) -> tuple[bool, str]:
+def is_correct_minerva(solution_str: str, gt: str, gt_need_extract: bool = False, answer_pattern: str = r"(?i)Answer\s*:\s*([^\n]+)") -> tuple[bool, str]:
     """Check if the solution is correct according to Minerva criteria.
 
     Args:
@@ -207,7 +205,15 @@ def is_correct_minerva(
     else:
         gt = normalize_final_answer(gt)
 
-    gt = str(int(float(gt)))  # in dapo, all answers are integers
+    try:
+        numeric_gt = Decimal(gt)
+    except InvalidOperation as exc:
+        raise ValueError(f"DAPO ground-truth label must be an integer, got {gt!r}") from exc
+
+    if not numeric_gt.is_finite() or numeric_gt != numeric_gt.to_integral_value():
+        raise ValueError(f"DAPO ground-truth label must be an integer, got {gt!r}")
+
+    gt = str(int(numeric_gt))
 
     return (pred == gt), pred
 
@@ -237,9 +243,7 @@ def is_correct_strict_box(pred: str, gt: str, pause_tokens_index: list[int] | No
     return 1 if (extracted_pred == gt) else -1, extracted_pred
 
 
-def verify(
-    solution_str: str, answer: str, strict_box_verify: bool = False, pause_tokens_index: list[int] | None = None
-) -> bool:
+def verify(solution_str: str, answer: str, strict_box_verify: bool = False, pause_tokens_index: list[int] | None = None) -> bool:
     """Verify if the solution is correct.
 
     Args:
diff --git a/tests/test_rm_math_dapo.py b/tests/test_rm_math_dapo.py
index 3a5e9a28b9..9d241b34ce 100644
--- a/tests/test_rm_math_dapo.py
+++ b/tests/test_rm_math_dapo.py
@@ -153,14 +153,37 @@ def test_is_correct_strict_box_no_box_returns_minus_one():
 
 @pytest.mark.unit
 def test_is_correct_minerva_matches_int_answer():
-    """Minerva path expects gt to coerce via ``int(float(gt))`` (line 210),
-    so floats / int-strings collapse to canonical int strings before
-    comparison."""
     correct, pred = is_correct_minerva("Long solution. Answer: 42", "42")
     assert correct is True
     assert pred == "42"
 
 
+@pytest.mark.unit
+@pytest.mark.parametrize("ground_truth", ["42.0", "4.2e1"])
+def test_is_correct_minerva_accepts_integral_numeric_labels(ground_truth):
+    correct, pred = is_correct_minerva("Answer: 42", ground_truth)
+    assert correct is True
+    assert pred == "42"
+
+
+@pytest.mark.unit
+def test_is_correct_minerva_preserves_large_integer_precision():
+    ground_truth = "9007199254740993"
+    correct, pred = is_correct_minerva(f"Answer: {ground_truth}", ground_truth)
+    assert correct is True
+    assert pred == ground_truth
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("ground_truth", ["0.5", "1/2", "1+2", "NaN", "Infinity"])
+def test_is_correct_minerva_rejects_non_integer_labels(ground_truth):
+    with pytest.raises(
+        ValueError,
+        match="DAPO ground-truth label must be an integer",
+    ):
+        is_correct_minerva("Answer: 0", ground_truth)
+
+
 @pytest.mark.unit
 def test_is_correct_minerva_takes_last_answer_match():
     """Multiple ``Answer:`` lines → ``re.findall`` returns the list and

From 5e345add05d4302608860da6cd30a4ffd0eda512 Mon Sep 17 00:00:00 2001
From: morluto <76467478+morluto@users.noreply.github.com>
Date: Tue, 14 Jul 2026 04:53:40 +0000
Subject: [PATCH 2/2] Apply repository formatting to DAPO scorer

---
 slime/rollout/rm_hub/math_dapo_utils.py | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/slime/rollout/rm_hub/math_dapo_utils.py b/slime/rollout/rm_hub/math_dapo_utils.py
index 35f58ab71c..1cab28ca0a 100644
--- a/slime/rollout/rm_hub/math_dapo_utils.py
+++ b/slime/rollout/rm_hub/math_dapo_utils.py
@@ -182,7 +182,9 @@ def normalize_final_answer(final_answer: str) -> str:
     return final_answer.strip()
 
 
-def is_correct_minerva(solution_str: str, gt: str, gt_need_extract: bool = False, answer_pattern: str = r"(?i)Answer\s*:\s*([^\n]+)") -> tuple[bool, str]:
+def is_correct_minerva(
+    solution_str: str, gt: str, gt_need_extract: bool = False, answer_pattern: str = r"(?i)Answer\s*:\s*([^\n]+)"
+) -> tuple[bool, str]:
     """Check if the solution is correct according to Minerva criteria.
 
     Args:
@@ -243,7 +245,9 @@ def is_correct_strict_box(pred: str, gt: str, pause_tokens_index: list[int] | No
     return 1 if (extracted_pred == gt) else -1, extracted_pred
 
 
-def verify(solution_str: str, answer: str, strict_box_verify: bool = False, pause_tokens_index: list[int] | None = None) -> bool:
+def verify(
+    solution_str: str, answer: str, strict_box_verify: bool = False, pause_tokens_index: list[int] | None = None
+) -> bool:
     """Verify if the solution is correct.
 
     Args: