test fix

Signed-off-by: Alexan <[email protected]>
NVIDIA · Nov 28, 2024 · 5821551 · 5821551
1 parent a3dc547
commit 5821551
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 1 deletion.
diff --git a/nemo_skills/evaluation/metrics/base.py b/nemo_skills/evaluation/metrics/base.py
@@ -59,7 +59,7 @@ def __init__(self, benchmark, extra_datasets=None, max_samples=-1, metric_class=
         if metric_class is None:
             # Setup metrics calculator
             benchmark_module, _ = get_dataset_module(benchmark, extra_datasets=extra_datasets)
-            self.metrics_calculator = benchmark_module.METRICS_CLASS()
+            self.metrics_calculator = benchmark_module.METRICS_CLASS
         else:
             self.metrics_calculator = metric_class
 

diff --git a/tests/test_generation.py b/tests/test_generation.py
@@ -101,6 +101,7 @@ def test_eval_mtbench_api():
         [f"/tmp/nemo-skills-tests/mtbench-api/eval-results/mt-bench/output.jsonl"],
     )["greedy"]
 
+    print(metrics)
     # not having other categories since we just ran with 2 samples
     assert metrics['average'] >= 7
     assert metrics['average_turn1'] >= 7