From a7164b620f862aeb4542713a6675ce3b4d9aea45 Mon Sep 17 00:00:00 2001 From: Lianmin Zheng Date: Sun, 17 Nov 2024 21:51:00 -0800 Subject: [PATCH] Tune the threshold for accuracy tests in CI (#2071) --- test/srt/test_eval_accuracy_large.py | 6 +++--- test/srt/test_moe_eval_accuracy_large.py | 6 +++--- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/test/srt/test_eval_accuracy_large.py b/test/srt/test_eval_accuracy_large.py index 000910cf2..22f7ab435 100644 --- a/test/srt/test_eval_accuracy_large.py +++ b/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__": diff --git a/test/srt/test_moe_eval_accuracy_large.py b/test/srt/test_moe_eval_accuracy_large.py index 401a47ce2..993e85a81 100644 --- a/test/srt/test_moe_eval_accuracy_large.py +++ b/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__":