fix(llmobs): handle None metadata in experiment dataset records

ddtrace/llmobs/_experiment.py

@@ -1649,7 +1649,7 @@ def as_dataframe(self) -> None:
                 flat_record[("expected_output", "")] = expected_output
                 column_tuples.add(("expected_output", ""))
 
-            metadata = record.get("metadata", {})
+            metadata = (record.get("metadata") or {})
             if isinstance(metadata, dict):
                 for metadata_col, metadata_val in metadata.items():
                     flat_record[("metadata", metadata_col)] = metadata_val
@@ -2000,7 +2000,7 @@ def _prepare_summary_evaluator_data(
             record: DatasetRecord = self._dataset[idx]
             inputs.append(record["input_data"])
             expected_outputs.append(record["expected_output"])
-            record_metadata = record.get("metadata", {})
+            record_metadata = (record.get("metadata") or {})
             metadata_list.append({**record_metadata, "experiment_config": self._config})
 
             eval_result_at_idx_by_name = eval_results[idx]["evaluations"]
@@ -2237,7 +2237,7 @@ async def _process_record(
                     tags["dataset_record_canonical_id"] = canonical_id
                 output_data = None
                 last_exc_info = None
-                record_metadata = record.get("metadata", {})
+                record_metadata = (record.get("metadata") or {})
                 task_args: list = [input_data, self._config]
                 if self._task_accepts_metadata:
                     task_args.append(record_metadata)
@@ -2358,7 +2358,7 @@ async def _evaluate_record(
         input_data = record["input_data"]
         output_data = task_result["output"]
         expected_output = record["expected_output"]
-        metadata = record.get("metadata", {})
+        metadata = (record.get("metadata") or {})
 
         async def _run_single_evaluator(
             evaluator: Union[EvaluatorType, AsyncEvaluatorType],

releasenotes/notes/fix-llmobs-experiment-none-metadata-7d4e2f9a8c1b3e5f.yaml

@@ -0,0 +1,9 @@
+---
+fixes:
+  - |
+    LLM Observability: Fixes a ``TypeError: 'NoneType' object is not a mapping`` raised
+    when running an experiment over dataset records whose ``metadata`` field is ``None``
+    (a common shape for records serialized with an explicit ``null``). Metadata reads in
+    ``ddtrace/llmobs/_experiment.py`` now fall back to an empty dict for both missing and
+    ``None`` values, so summary evaluators, task argument plumbing, and per-record evaluator
+    context no longer crash on these records.

tests/llmobs/test_experiments.py

@@ -1753,6 +1753,45 @@ def test_experiment_run_summary_evaluators(llmobs, test_dataset_one_record):
     }
 
 
+def test_experiment_run_summary_evaluators_handles_none_metadata(llmobs):
+    """Regression: records whose `metadata` field is explicitly None must not crash
+    the summary evaluator path. `dict.get("metadata", {})` returns the stored None
+    (not the default), which then fails when spread into a dict literal. See
+    _prepare_summary_evaluator_data and the per-record paths that build
+    `{**metadata, "experiment_config": ...}`.
+    """
+    dataset = Dataset(
+        name="test_dataset_none_metadata",
+        project={"name": "test_project", "_id": "proj_123"},
+        dataset_id="ds_none_metadata",
+        records=[
+            {
+                "record_id": "rec_1",
+                "input_data": {"prompt": "What is the capital of France?"},
+                "expected_output": {"answer": "Paris"},
+                "metadata": None,
+            }
+        ],
+        description="Dataset whose records carry metadata=None",
+        latest_version=1,
+        version=1,
+        _dne_client=None,
+    )
+    exp = llmobs.experiment(
+        "test_experiment",
+        dummy_task,
+        dataset,
+        [dummy_evaluator],
+        summary_evaluators=[dummy_summary_evaluator],
+    )
+    task_results = asyncio.run(exp._experiment._run_task(1, run=run_info_with_stable_id(0), raise_errors=False))
+    eval_results = asyncio.run(exp._experiment._run_evaluators(task_results, raise_errors=False))
+    summary_eval_results = asyncio.run(
+        exp._experiment._run_summary_evaluators(task_results, eval_results, raise_errors=True)
+    )
+    assert summary_eval_results[0]["evaluations"]["dummy_summary_evaluator"]["error"] is None
+
+
 def test_experiment_run_evaluators_error(llmobs, test_dataset_one_record):
     exp = llmobs.experiment("test_experiment", dummy_task, test_dataset_one_record, [faulty_evaluator])
     task_results = asyncio.run(exp._experiment._run_task(1, run=run_info_with_stable_id(0), raise_errors=False))