diff --git a/tests/test_bigcode_eval.py b/tests/test_bigcode_eval.py index afd1355a..6df8e5c7 100644 --- a/tests/test_bigcode_eval.py +++ b/tests/test_bigcode_eval.py @@ -30,13 +30,13 @@ def test_lm_eval(self): user_model=user_model, tokenizer=tokenizer, tasks="humaneval", - n_samples=20, - batch_size=10, + n_samples=2, + batch_size=2, allow_code_execution=True, - limit=20, + limit=10, ) results = evaluate(args) - self.assertEqual(results["humaneval"]["pass@1"], 0.05) + self.assertEqual(results["humaneval"]["pass@1"], 0.1) if __name__ == "__main__":