Typos

sparse_autoencoder/autoencoder/types.py

@@ -8,7 +8,7 @@ class ResetOptimizerParameterDetails(NamedTuple):
     """Reset Optimizer Parameter Details.
 
     Details of a parameter that should be reset in the optimizer, when resetting
-    it's corresponding dictionary vectors.
+    its corresponding dictionary vectors.
     """
 
     parameter: Parameter

sparse_autoencoder/loss/decoded_activations_l2.py

@@ -14,7 +14,7 @@ class L2ReconstructionLoss(AbstractLoss):
     """L2 Reconstruction loss.
 
     L2 reconstruction loss is calculated as the sum squared error between each each input vector
-    and it's corresponding decoded vector. The original paper found that models trained with some
+    and its corresponding decoded vector. The original paper found that models trained with some
     loss functions such as cross-entropy loss generally prefer to represent features
     polysemantically, whereas models trained with L2 may achieve the same loss for both
     polysemantic and monosemantic representations of true features.

sparse_autoencoder/source_data/text_dataset.py

@@ -151,7 +151,7 @@ def push_to_hugging_face_hub(
                 is set.
             n_shards: Number of shards to split the dataset into. A high number is recommended
                 here to allow for flexible distributed training of SAEs across nodes (where e.g.
-                each node fetches it's own shard).
+                each node fetches its own shard).
             revision: Branch to push to.
             private: Whether to save the dataset privately.
 

sparse_autoencoder/train/utils/wandb_sweep_types.py

@@ -341,7 +341,7 @@ def __repr__(self) -> str:
 
 
 @dataclass(frozen=True)
-class NestedParameter(ABC):  # noqa: B024 (abstract so that we can check against it's type)
+class NestedParameter(ABC):  # noqa: B024 (abstract so that we can check against its type)
     """Nested Parameter.
 
     Example: