refactor: deduplicate trainer and clean up load_data API

src/bakery/cli.py

@@ -245,15 +245,8 @@ def main():
     if data_config.eval_dataset_split and data_config.dataset:
         print(f"  Loading eval split: {data_config.eval_dataset_split}")
         eval_prompts, eval_responses = load_data(
-            type(
-                "_DC",
-                (),
-                {
-                    "dataset": data_config.dataset,
-                    "dataset_split": data_config.eval_dataset_split,
-                    "training_prompts": None,
-                },
-            )()
+            dataset=data_config.dataset,
+            dataset_split=data_config.eval_dataset_split,
         )
         eval_dataset = create_dataset(eval_prompts, eval_responses)
         print(f"  Eval samples: {len(eval_dataset)}")

src/bakery/data.py

@@ -91,19 +91,35 @@ def build_system_prompt(
 
 
 def load_data(
-    data_config: DataConfig,
+    data_config: Optional[DataConfig] = None,
+    *,
+    dataset: Optional[str] = None,
+    dataset_split: str = "train",
+    training_prompts: Optional[List[str]] = None,
 ) -> Tuple[List[str], Optional[List[str]]]:
-    """Load training data from config.
+    """Load training data from config or explicit parameters.
+
+    Can be called with a DataConfig object (backward-compatible) or with
+    explicit keyword arguments for dataset/split/training_prompts.
 
     Returns:
         (prompts, responses) where responses is None if only prompts are
         available (triggering on-the-fly trajectory generation).
     """
-    if data_config.dataset:
-        return load_dataset(data_config.dataset, data_config.dataset_split)
+    if data_config is not None:
+        dataset = dataset or data_config.dataset
+        dataset_split = (
+            data_config.dataset_split
+            if dataset == data_config.dataset
+            else dataset_split
+        )
+        training_prompts = training_prompts or data_config.training_prompts
+
+    if dataset:
+        return load_dataset(dataset, dataset_split)
 
-    if data_config.training_prompts:
-        return data_config.training_prompts, None
+    if training_prompts:
+        return training_prompts, None
 
     raise ValueError(
         "No training data configured. Provide 'dataset' or 'training_prompts'."