Tune the threshold for accuracy tests in CI (#2071)

test/srt/test_eval_accuracy_large.py

@@ -42,7 +42,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.71, f"{metrics}"
+        self.assertGreater(metrics["score"], 0.71)
 
     def test_human_eval(self):
         args = SimpleNamespace(
@@ -54,7 +54,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.64, f"{metrics}"
+        self.assertGreater(metrics["score"], 0.64)
 
     def test_mgsm_en(self):
         args = SimpleNamespace(
@@ -66,7 +66,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.84, f"{metrics}"
+        self.assertGreater(metrics["score"], 0.84)
 
 
 if __name__ == "__main__":

test/srt/test_moe_eval_accuracy_large.py

@@ -47,7 +47,7 @@ class TestMoEEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.62, f"{metrics}"
+        self.assertGreater(metrics["score"], 0.62)
 
     def test_human_eval(self):
         args = SimpleNamespace(
@@ -59,7 +59,7 @@ class TestMoEEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.42, f"{metrics}"
+        self.assertGreater(metrics["score"], 0.41)
 
     def test_mgsm_en(self):
         args = SimpleNamespace(
@@ -71,7 +71,7 @@ class TestMoEEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.62, f"{metrics}"
+        self.assertGreater(metrics["score"], 0.62)
 
 
 if __name__ == "__main__":