feat: Add Language Adaptation Transforms for Cross-Lingual AIRT Probing

dreadnode/eval/hooks/transforms.py

@@ -35,7 +35,7 @@ async def hook(event: "EvalEvent") -> "EvalReaction | None":  # noqa: PLR0911
             if create_task:
                 from dreadnode import task as dn_task
 
-                task_kwargs = event.task_kwargs
+                input_data = event.task_kwargs
 
                 @dn_task(
                     name=f"transform - input ({len(transforms)} transforms)",
@@ -44,11 +44,11 @@ async def hook(event: "EvalEvent") -> "EvalReaction | None":  # noqa: PLR0911
                     log_output=True,
                 )
                 async def apply_task(
-                    data: dict[str, t.Any] = task_kwargs,  # Use extracted variable
+                    data: dict[str, t.Any],
                 ) -> dict[str, t.Any]:
                     return await apply_transforms_to_kwargs(data, transforms)
 
-                transformed = await apply_task()
+                transformed = await apply_task(input_data)
                 return ModifyInput(task_kwargs=transformed)
 
             # Direct application
@@ -73,10 +73,12 @@ async def apply_task(
                 log_inputs=True,
                 log_output=True,
             )
-            async def apply_task(data: t.Any = output_data) -> t.Any:  # Use extracted variable
+            async def apply_task(
+                data: t.Any,
+            ) -> t.Any:
                 return await apply_transforms_to_value(data, transforms)
 
-            transformed = await apply_task()
+            transformed = await apply_task(output_data)
             return ModifyOutput(output=transformed)
 
         # Direct application

dreadnode/optimization/study.py

@@ -8,6 +8,7 @@
 from pydantic import ConfigDict, Field, FilePath, SkipValidation, computed_field
 
 from dreadnode.common_types import AnyDict
+from dreadnode.data_types.message import Message
 from dreadnode.error import AssertionFailedError
 from dreadnode.eval import InputDataset
 from dreadnode.eval.eval import Eval
@@ -379,17 +380,16 @@ async def _run_evaluation(
         )
         logger.trace(f"Candidate: {trial.candidate!r}")
 
-        # if dataset == [{}] or (isinstance(dataset, list) and len(dataset) == 1 and not dataset[0]):
-        #     # Dataset is empty - this is a Study/Attack where the candidate IS the input
-        #     dataset = [{"message": trial.candidate}]
-        #     dataset_input_mapping = ["message"]
-        # else:
-        #     dataset_input_mapping = None
+        dataset_input_mapping = None
+        # If dataset is empty and candidate is a Message, this is an airt attack scenario
+        if dataset == [{}] and isinstance(trial.candidate, Message):
+            dataset = [{"message": trial.candidate}]
+            dataset_input_mapping = ["message"]
 
         evaluator = Eval(
             task=task,
             dataset=dataset,
-            # dataset_input_mapping=dataset_input_mapping,
+            dataset_input_mapping=dataset_input_mapping,
             scorers=scorers,
             hooks=self.hooks,
             max_consecutive_errors=self.max_consecutive_errors,