Fixing issues with multichoice_continuations_start_space - was not pa…

…rsed properly (#232)

* fixing  issues - was not parsed properly

* Update src/lighteval/models/model_config.py

Co-authored-by: Nathan Habib <[email protected]>

* fix

---------

Co-authored-by: Nathan Habib <[email protected]>

examples/model_configs/base_model.yaml

@@ -9,5 +9,4 @@ model:
     adapter_weights: false # set to True of your model has been trained with peft, also need to provide the base model name
     base_model: null # path to the base_model
   generation:
-    multichoice_continuations_start_space: false # Whether to force multiple choice continuations to start with a space
-    no_multichoice_continuations_start_space: false # Whether to force multiple choice continuations to not start with a space
+    multichoice_continuations_start_space: null # If true/false, will force multiple choice continuations to start/not start with a space. If none, will do nothing

examples/nanotron/lighteval_config_override_template.yaml

@@ -24,7 +24,6 @@ lighteval:
     dataset_loading_processes: 8
     max_samples: 10
     multichoice_continuations_start_space: null
-    no_multichoice_continuations_start_space: null
     num_fewshot_seeds: null
     tasks: early-signal
     # tasks: custom|hellaswag|0

src/lighteval/models/base_model.py

@@ -296,10 +296,11 @@ def _check_continuations_start_space(self, continuation: str) -> str:
         - None (Don't touch - default)
         Todo: find a way to add this back WITHOUT breaking compatibility with the harness
         """
-        if self.multichoice_continuations_start_space is True and continuation[0] != " ":
-            continuation = " " + continuation
-        if self.multichoice_continuations_start_space is False and continuation[0] == " ":
-            continuation = continuation.lstrip()
+        if self.multichoice_continuations_start_space is not None:
+            if self.multichoice_continuations_start_space and continuation[0] != " ":
+                continuation = " " + continuation
+            if not self.multichoice_continuations_start_space and continuation[0] == " ":
+                continuation = continuation.lstrip()
         return continuation
 
     def _model_call(self, inputs: torch.Tensor) -> torch.Tensor:
@@ -949,7 +950,7 @@ def loglikelihood_single_token(
             if any(len(c) > 1 for c in continuations_enc):
                 raise ValueError(
                     f"Trying to do single token multiple choice but one choice has several tokens: {continuations_enc}. "
-                    "If the additional pre-token is a space, try to set --no_multichoice_continuations_start_space "
+                    "If the additional pre-token is a space, try to set `multichoice_continuations_start_space=False` in the model parameters "
                 )
             request.tokenized_continuation = continuations_enc
 

src/lighteval/models/model_config.py

@@ -34,6 +34,7 @@
     NO_AUTOGPTQ_ERROR_MSG,
     NO_BNB_ERROR_MSG,
     NO_PEFT_ERROR_MSG,
+    boolstring_to_bool,
     is_accelerate_available,
     is_autogptq_available,
     is_bnb_available,
@@ -76,6 +77,7 @@ class BaseModelConfig:
             space at the start of each continuation in multichoice generation.
             For example, context: "What is the capital of France?" and choices: "Paris", "London".
             Will be tokenized as: "What is the capital of France? Paris" and "What is the capital of France? London".
+            True adds a space, False strips a space, None does nothing
         subfolder (Optional[str]): The subfolder within the model repository.
         revision (str): The revision of the model.
         batch_size (int): The batch size for model training.
@@ -127,6 +129,9 @@ class BaseModelConfig:
     compile: bool = False
 
     def __post_init__(self):
+        # Making sure this parameter is a boolean
+        self.multichoice_continuations_start_space = boolstring_to_bool(self.multichoice_continuations_start_space)
+
         if self.quantization_config is not None and not is_bnb_available():
             raise ImportError(NO_BNB_ERROR_MSG)
 
@@ -350,15 +355,21 @@ def create_model_config(  # noqa: C901
         return InferenceModelConfig(model=config["base_params"]["endpoint_name"])
 
     if config["type"] == "base":
-        # Tests on the multichoice space parameters
-        multichoice_continuations_start_space = config["generation"]["multichoice_continuations_start_space"]
-        no_multichoice_continuations_start_space = config["generation"]["no_multichoice_continuations_start_space"]
-        if not multichoice_continuations_start_space and not no_multichoice_continuations_start_space:
-            multichoice_continuations_start_space = None
-        if multichoice_continuations_start_space and no_multichoice_continuations_start_space:
-            raise ValueError(
-                "You cannot force both the multichoice continuations to start with a space and not to start with a space"
-            )
+        # Creating the multichoice space parameters
+        # We need to take into account possible conversion issues from our different input formats
+        multichoice_continuations_start_space = boolstring_to_bool(
+            config["generation"]["multichoice_continuations_start_space"]
+        )
+
+        if multichoice_continuations_start_space is not None:
+            if multichoice_continuations_start_space:
+                hlog(
+                    "You set `multichoice_continuations_start_space` to true. This will force multichoice continuations to use a starting space"
+                )
+            else:
+                hlog(
+                    "You set `multichoice_continuations_start_space` to false. This will remove a leading space from multichoice continuations, if present."
+                )
 
         # Creating optional quantization configuration
         if config["base_params"]["dtype"] == "4bit":

src/lighteval/models/nanotron_model.py

@@ -63,7 +63,7 @@
     LoglikelihoodRequest,
     LoglikelihoodRollingRequest,
 )
-from lighteval.utils import as_list
+from lighteval.utils import as_list, boolstring_to_bool
 from lighteval.utils_parallelism import find_executable_batch_size
 
 
@@ -113,20 +113,10 @@ def __init__(
             # To implement PP parallelism we need to think about how we want to sync the output for the PP ranks without outputs
             raise ValueError("PP parallelism is not supported yet")
 
-        multichoice_continuations_start_space = lighteval_config.tasks.multichoice_continuations_start_space
-        if (
-            not multichoice_continuations_start_space
-            and not lighteval_config.tasks.no_multichoice_continuations_start_space
-        ):
-            multichoice_continuations_start_space = None
-            # multichoice_continuations_start_space can be True (forcing space), False (forcing no space) or None (no forcing)
-        if (
+        # multichoice_continuations_start_space can be True (forcing space), False (forcing no space) or None (no forcing)
+        multichoice_continuations_start_space = boolstring_to_bool(
             lighteval_config.tasks.multichoice_continuations_start_space
-            and lighteval_config.tasks.no_multichoice_continuations_start_space
-        ):
-            raise ValueError(
-                "You cannot force both the multichoice continuations to start with a space and not to start with a space"
-            )
+        )
 
         self.generation_config = lighteval_config.generation
         if isinstance(self.generation_config, dict):
@@ -409,10 +399,11 @@ def _check_continuations_start_space(self, continuation: str) -> str:
         - None (Don't touch - default)
         Todo: find a way to add this back WITHOUT breaking compatibility with the harness
         """
-        if self.multichoice_continuations_start_space is True and continuation[0] != " ":
-            continuation = " " + continuation
-        if self.multichoice_continuations_start_space is False and continuation[0] == " ":
-            continuation = continuation.lstrip()
+        if self.multichoice_continuations_start_space is not None:
+            if self.multichoice_continuations_start_space and continuation[0] != " ":
+                continuation = " " + continuation
+            if not self.multichoice_continuations_start_space and continuation[0] == " ":
+                continuation = continuation.lstrip()
         return continuation
 
     def loglikelihood_single_token(
@@ -443,7 +434,7 @@ def loglikelihood_single_token(
             if any(len(c) > 1 for c in continuations_enc):
                 raise ValueError(
                     f"Trying to do single token multiple choice but one choice has several tokens: {continuations_enc}. "
-                    "If the additional pre-token is a space, try to set --no_multichoice_continuations_start_space "
+                    "If the additional pre-token is a space, try to set multichoice_continuations_start_space=False in the model parameters "
                 )
             request.tokenized_continuation = continuations_enc
 

src/lighteval/utils.py

@@ -145,6 +145,24 @@ def flatten(item: list[Union[list, str]]) -> list[str]:
     return flat_item
 
 
+def boolstring_to_bool(x: Union[str, bool, int]) -> Union[bool, None]:
+    """Allows to manage string or bool to bool conversion, in case a configuration input is badly formatted.
+
+    Args:
+        x (str): A string (true, false, True, False, ...)
+
+    Returns:
+        Union[bool, None]: the corresponding boolean
+    """
+    if x in [None, "None", "null", ""]:
+        return None
+    if x in ["True", "true", True, 1]:
+        return True
+    if x in ["False", "false", False, 0]:
+        return False
+    raise ValueError(f"You tried to convert {x} to a boolean but it's not possible.")
+
+
 def is_accelerate_available() -> bool:
     return importlib.util.find_spec("accelerate") is not None