From 847c4a8a048e527df97d319645e7d577c5b2938b Mon Sep 17 00:00:00 2001
From: Surbhi Jain <sjsurbhi@google.com>
Date: Wed, 29 Apr 2026 18:33:46 +0000
Subject: [PATCH] Add unit tests for evaluate_rl.py

---
 tests/post_training/unit/evaluate_rl_test.py | 93 ++++++++++++++++++++
 1 file changed, 93 insertions(+)

diff --git a/tests/post_training/unit/evaluate_rl_test.py b/tests/post_training/unit/evaluate_rl_test.py
index 205e4a8797..a28b0cda35 100644
--- a/tests/post_training/unit/evaluate_rl_test.py
+++ b/tests/post_training/unit/evaluate_rl_test.py
@@ -14,9 +14,11 @@
 
 """Unit tests for evaluate_rl.py (CPU-only)."""
 
+import json
 import unittest
 import pytest
 from types import SimpleNamespace
+from unittest import mock
 
 from maxtext.trainers.post_train.rl import evaluate_rl
 
@@ -208,5 +210,96 @@ def test_pass_at_1_all_correct(self):
     self.assertAlmostEqual(has_correct_format, 1.0)
 
 
+class TestEvaluate(unittest.TestCase):
+  """Tests for the main evaluate() function."""
+
+  def setUp(self):
+    self.config = _make_config()
+    self.mock_cluster = mock.Mock()
+    self.dataset = [
+        {
+            "question": ["q1", "q2"],
+            "answer": [json.dumps(["a1"]), json.dumps(["a2"])],
+            "prompts": ["p1", "p2"],
+        }
+    ]
+
+  @pytest.mark.cpu_only
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.generate_responses")
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.score_responses")
+  def test_evaluate_pass_mode(self, mock_score, mock_generate):
+    """Test basic evaluation flow in 'pass' mode."""
+    config = _make_config(eval_mode="pass")
+    # Two items in batch: one correct, one incorrect.
+    mock_generate.return_value = [["r1"], ["r2"]]
+    mock_score.side_effect = [
+        (True, True, True),  # First item is correct
+        (False, False, False),  # Second item is incorrect
+    ]
+
+    stats, response_list = evaluate_rl.evaluate(config, self.dataset, self.mock_cluster)
+
+    self.assertEqual(stats, (1, 2, 50.0, 50.0, 50.0))
+    self.assertEqual(response_list, [])
+    self.assertEqual(mock_generate.call_count, 1)
+    self.assertEqual(mock_score.call_count, 2)
+
+  @pytest.mark.cpu_only
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.generate_responses")
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.score_responses")
+  def test_evaluate_pass_at_1_mode(self, mock_score, mock_generate):
+    """Test evaluation flow in 'pass_at_1' mode, which returns floats."""
+    config = _make_config(eval_mode="pass_at_1")
+    mock_generate.return_value = [["r1a", "r1b"], ["r2a", "r2b"]]
+    # First item: 1/2 correct (0.5)
+    # Second item: 2/2 correct (1.0)
+    mock_score.side_effect = [
+        (0.5, 0.5, 1.0),
+        (1.0, 1.0, 1.0),
+    ]
+
+    stats, _ = evaluate_rl.evaluate(config, self.dataset, self.mock_cluster)
+
+    # Total correct = 0.5 + 1.0 = 1.5
+    # Total items = 2
+    # Accuracy = (1.5 / 2) * 100 = 75.0
+    self.assertEqual(stats, (1.5, 2, 75.0, 75.0, 100.0))
+
+  @pytest.mark.cpu_only
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.generate_responses")
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.score_responses")
+  def test_make_lst_correct(self, mock_score, mock_generate):
+    """Test that make_lst=True, corr_lst=True returns only correct items."""
+    mock_generate.return_value = [["r1"], ["r2"]]
+    mock_score.side_effect = [(True, True, True), (False, False, False)]
+
+    _, response_list = evaluate_rl.evaluate(self.config, self.dataset, self.mock_cluster, make_lst=True, corr_lst=True)
+
+    self.assertEqual(len(response_list), 1)
+    self.assertEqual(response_list[0], ("q1", ["a1"], ["r1"]))
+
+  @pytest.mark.cpu_only
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.generate_responses")
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.score_responses")
+  def test_make_lst_incorrect(self, mock_score, mock_generate):
+    """Test that make_lst=True, corr_lst=False returns only incorrect items."""
+    mock_generate.return_value = [["r1"], ["r2"]]
+    mock_score.side_effect = [(True, True, True), (False, False, False)]
+
+    _, response_list = evaluate_rl.evaluate(self.config, self.dataset, self.mock_cluster, make_lst=True, corr_lst=False)
+
+    self.assertEqual(len(response_list), 1)
+    self.assertEqual(response_list[0], ("q2", ["a2"], ["r2"]))
+
+  @pytest.mark.cpu_only
+  @mock.patch("maxtext.trainers.post_train.rl.evaluate_rl.generate_responses")
+  def test_empty_dataset(self, mock_generate):
+    """Test that evaluation on an empty dataset returns zero stats."""
+    stats, response_list = evaluate_rl.evaluate(self.config, [], self.mock_cluster)
+    self.assertEqual(stats, (0, 0, 0, 0, 0))
+    self.assertEqual(response_list, [])
+    mock_generate.assert_not_called()
+
+
 if __name__ == "__main__":
   unittest.main()