Update BenchmarkScoreEncoder to handle serialization errors

src/modelbench/record.py

@@ -116,4 +116,7 @@ def default(self, o):
         elif isinstance(o, datetime):
             return str(o)
         else:
-            return super().default(o)
+            try:
+                return super().default(o)
+            except TypeError as e:
+                return str(e)

tests/modelbench_tests/test_record.py

@@ -10,11 +10,15 @@
 
 from modelbench.benchmarks import BenchmarkScore, GeneralPurposeAiChatBenchmarkV1, SecurityBenchmark, SecurityScore
 from modelbench.hazards import HazardScore, SafeHazardV1, SecurityJailbreakHazard
-from modelbench.record import benchmark_code_info, BenchmarkScoreEncoder, dump_json
+from modelbench.record import BenchmarkScoreEncoder, benchmark_code_info, dump_json
 from modelbench.scoring import ValueEstimate
 from modelbench.standards import Standards
+from modelgauge.auth.openai_compatible_secrets import OpenAIApiKey
 from modelgauge.locales import EN_US
 from modelgauge.record_init import InitializationRecord
+from modelgauge.sut import PromptResponseSUT
+from modelgauge.sut_decorator import modelgauge_sut
+from modelgauge.suts.openai_client import OpenAIChat
 
 
 def benchmark_run_record(benchmark_score):
@@ -30,8 +34,7 @@ def secrets():
     return {"together": {"api_key": "fake"}, "modellab_files": {"token": "fake"}}
 
 
-@pytest.fixture()
-def benchmark_score(end_time, sut):
+def benchmark_score_gen(end_time, sut):
     bd = GeneralPurposeAiChatBenchmarkV1(EN_US, "practice")
     low_est = ValueEstimate.make(0.5, 10)
     high_est = ValueEstimate.make(0.8, 20)
@@ -61,6 +64,32 @@ def benchmark_score(end_time, sut):
     return bs
 
 
+@pytest.fixture()
+def benchmark_score(end_time, sut):
+    return benchmark_score_gen(end_time, sut)
+
+
+@pytest.fixture()
+def unserializable_sut():
+    @modelgauge_sut(capabilities=[])
+    class UnserializableSUT(PromptResponseSUT):
+        def __init__(self, uid, unserializable_init_param):
+            super().__init__(uid)
+
+        def evaluate(self, request):
+            pass
+
+        def translate_response(self, request, response):
+            pass
+
+    return UnserializableSUT("unserializable_sut", unserializable_init_param=lambda x: x)
+
+
+@pytest.fixture()
+def benchmark_score_with_unserializable_sut(end_time, unserializable_sut):
+    return benchmark_score_gen(end_time, unserializable_sut)
+
+
 @pytest.fixture()
 def security_score(monkeypatch, tmp_path, end_time, sut):
     # Patch the standards in case the benchmark is not yet calibrated.
@@ -121,6 +150,11 @@ def test_sut(sut):
     assert "initialization" in encoded
 
 
+def test_unserializable_sut_within_benchmark_score(benchmark_score_with_unserializable_sut):
+    encoded = encode(benchmark_score_with_unserializable_sut)
+    assert '"unserializable_init_param": "Object of type function is not JSON serializable"' in encoded
+
+
 def test_value_estimate():
     ve = ValueEstimate.make(0.5, 1000)
     j = encode_and_parse(ve)