feat(data-process): 支持思维链输出类型

backend/tests/test_data_process_store.py

@@ -15,6 +15,7 @@ from app.modules.data_process.store import (
     InvalidStateError,
     _decode_row,
     _preview_config_changed,
+    _reasoning_output_is_valid,
     _source_storage_descriptor,
 )
 
@@ -114,14 +115,25 @@ class _PublishConnection:
             )
         if normalized.startswith("INSERT INTO dataset_records"):
             self.records.append(
-                {"dataset_id": params[1], "line_no": params[4], "split": params[5]}
+                {
+                    "dataset_id": params[1],
+                    "line_no": params[4],
+                    "split": params[5],
+                    "output": params[8],
+                    "raw": json.loads(params[9]),
+                }
             )
         return _Result()
 
 
 class _PublishStore(DataProcessStore):
-    def __init__(self, conn: _PublishConnection):
+    def __init__(
+        self,
+        conn: _PublishConnection,
+        task_config: dict[str, Any] | None = None,
+    ):
         self._conn = conn
+        self._task_config = task_config or {}
 
     @contextmanager
     def connect(self) -> Iterator[_PublishConnection]:
@@ -135,7 +147,7 @@ class _PublishStore(DataProcessStore):
             "id": task_id,
             "status": "completed",
             "description": "",
-            "config": {},
+            "config": self._task_config,
             "output_dataset_id": train_dataset and train_dataset["id"],
         }
 
@@ -967,7 +979,11 @@ def test_publish_creates_three_independent_datasets_with_exact_counts() -> None:
             "status": "valid",
             "instruction": f"问题 {index}",
             "input": "",
-            "output": f"答案 {index}",
+            "output": (
+                "<think>\n先读取制度条款。\n</think>\n答案 0"
+                if index == 0
+                else f"答案 {index}"
+            ),
             "preview_item_id": f"preview-{index}",
         }
         for index in range(28)
@@ -993,6 +1009,10 @@ def test_publish_creates_three_independent_datasets_with_exact_counts() -> None:
         item["id"] for item in conn.datasets
     }
     assert len(conn.records) == 28
+    reasoning_record = next(
+        item for item in conn.records if item["output"].startswith("<think>")
+    )
+    assert reasoning_record["raw"]["output"] == reasoning_record["output"]
     assert published["dataset"]["type"] == "train"
     assert len(published["datasets"]) == 3
     assert published["split_counts"] == {"train": 22, "validation": 3, "test": 3}
@@ -1078,6 +1098,49 @@ def test_publish_keeps_all_three_datasets_when_a_small_split_is_empty() -> None:
     assert len(conn.files) == 3
 
 
+@pytest.mark.parametrize(
+    ("output", "expected"),
+    [
+        ("<think>\n推理步骤\n</think>\n最终答案", True),
+        ("<think></think>\n最终答案", False),
+        ("<think>只有推理</think>", False),
+        ("没有标签的最终答案", False),
+        ("<think>外层<think>嵌套</think></think>答案", False),
+    ],
+)
+def test_reasoning_output_validator_requires_one_complete_pair(
+    output: str,
+    expected: bool,
+) -> None:
+    assert _reasoning_output_is_valid(output) is expected
+
+
+def test_publish_rejects_invalid_reasoning_output_format() -> None:
+    conn = _PublishConnection(
+        [
+            {
+                "id": "result-reasoning-invalid",
+                "status": "valid",
+                "instruction": "需要推理的问题",
+                "input": "",
+                "output": "只有最终答案",
+                "preview_item_id": "preview-reasoning-invalid",
+            }
+        ]
+    )
+
+    with pytest.raises(InvalidStateError, match="1 invalid results"):
+        _PublishStore(conn, {"output_type": "reasoning"}).publish(
+            "task-reasoning-invalid",
+            {
+                "dataset_name": "无效思维链",
+                "storage_type": "local",
+                "format": "alpaca_jsonl",
+                "split": {"train": 80, "validation": 10, "test": 10},
+            },
+        )
+
+
 def test_source_storage_descriptor_accepts_owned_local_and_legacy_db_references() -> None:
     task_id = "dpt_task"
     source_file_id = "dpsf_source"