sae-lens 6.20.1__tar.gz → 6.22.2__tar.gz

{sae_lens-6.20.1 → sae_lens-6.22.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sae-lens
-Version: 6.20.1
+Version: 6.22.2
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 License-File: LICENSE

{sae_lens-6.20.1 → sae_lens-6.22.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "sae-lens"
-version = "6.20.1"
+version = "6.22.2"
 description = "Training and Analyzing Sparse Autoencoders (SAEs)"
 authors = ["Joseph Bloom"]
 readme = "README.md"

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.20.1"
+__version__ = "6.22.2"
 
 import logging
 

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/analysis/hooked_sae_transformer.py

@@ -3,7 +3,6 @@ from contextlib import contextmanager
 from typing import Any, Callable
 
 import torch
-from jaxtyping import Float
 from transformer_lens.ActivationCache import ActivationCache
 from transformer_lens.components.mlps.can_be_used_as_mlp import CanBeUsedAsMLP
 from transformer_lens.hook_points import HookPoint  # Hooking utilities
@@ -11,8 +10,8 @@ from transformer_lens.HookedTransformer import HookedTransformer
 
 from sae_lens.saes.sae import SAE
 
-SingleLoss = Float[torch.Tensor, ""]  # Type alias for a single element tensor
-LossPerToken = Float[torch.Tensor, "batch pos-1"]
+SingleLoss = torch.Tensor  # Type alias for a single element tensor
+LossPerToken = torch.Tensor
 Loss = SingleLoss | LossPerToken
 
 
@@ -171,12 +170,7 @@ class HookedSAETransformer(HookedTransformer):
         reset_saes_end: bool = True,
         use_error_term: bool | None = None,
         **model_kwargs: Any,
-    ) -> (
-        None
-        | Float[torch.Tensor, "batch pos d_vocab"]
-        | Loss
-        | tuple[Float[torch.Tensor, "batch pos d_vocab"], Loss]
-    ):
+    ) -> None | torch.Tensor | Loss | tuple[torch.Tensor, Loss]:
         """Wrapper around HookedTransformer forward pass.
 
         Runs the model with the given SAEs attached for one forward pass, then removes them. By default, will reset all SAEs to original state after.
@@ -203,10 +197,7 @@ class HookedSAETransformer(HookedTransformer):
         remove_batch_dim: bool = False,
         **kwargs: Any,
     ) -> tuple[
-        None
-        | Float[torch.Tensor, "batch pos d_vocab"]
-        | Loss
-        | tuple[Float[torch.Tensor, "batch pos d_vocab"], Loss],
+        None | torch.Tensor | Loss | tuple[torch.Tensor, Loss],
         ActivationCache | dict[str, torch.Tensor],
     ]:
         """Wrapper around 'run_with_cache' in HookedTransformer.

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/cache_activations_runner.py

@@ -9,7 +9,6 @@ import torch
 from datasets import Array2D, Dataset, Features, Sequence, Value
 from datasets.fingerprint import generate_fingerprint
 from huggingface_hub import HfApi
-from jaxtyping import Float, Int
 from tqdm.auto import tqdm
 from transformer_lens.HookedTransformer import HookedRootModule
 
@@ -318,8 +317,8 @@ class CacheActivationsRunner:
     def _create_shard(
         self,
         buffer: tuple[
-            Float[torch.Tensor, "(bs context_size) d_in"],
-            Int[torch.Tensor, "(bs context_size)"] | None,
+            torch.Tensor,  # shape: (bs context_size) d_in
+            torch.Tensor | None,  # shape: (bs context_size) or None
         ],
     ) -> Dataset:
         hook_names = [self.cfg.hook_name]

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/pretrained_saes.yaml

@@ -14916,6 +14916,30 @@ qwen2.5-7b-instruct-andyrdt:
     path: resid_post_layer_27/trainer_1
     neuronpedia: qwen2.5-7b-it/27-resid-post-aa
 
+gpt-oss-20b-andyrdt:
+  conversion_func: dictionary_learning_1
+  model: openai/gpt-oss-20b
+  repo_id: andyrdt/saes-gpt-oss-20b
+  saes:
+  - id: resid_post_layer_3_trainer_0
+    path: resid_post_layer_3/trainer_0
+    neuronpedia: gpt-oss-20b/3-resid-post-aa
+  - id: resid_post_layer_7_trainer_0
+    path: resid_post_layer_7/trainer_0
+    neuronpedia: gpt-oss-20b/7-resid-post-aa
+  - id: resid_post_layer_11_trainer_0
+    path: resid_post_layer_11/trainer_0
+    neuronpedia: gpt-oss-20b/11-resid-post-aa
+  - id: resid_post_layer_15_trainer_0
+    path: resid_post_layer_15/trainer_0
+    neuronpedia: gpt-oss-20b/15-resid-post-aa
+  - id: resid_post_layer_19_trainer_0
+    path: resid_post_layer_19/trainer_0
+    neuronpedia: gpt-oss-20b/19-resid-post-aa
+  - id: resid_post_layer_23_trainer_0
+    path: resid_post_layer_23/trainer_0
+    neuronpedia: gpt-oss-20b/23-resid-post-aa
+
 goodfire-llama-3.3-70b-instruct:
   conversion_func: goodfire
   model: meta-llama/Llama-3.3-70B-Instruct
@@ -14924,6 +14948,7 @@ goodfire-llama-3.3-70b-instruct:
   - id: layer_50
     path: Llama-3.3-70B-Instruct-SAE-l50.pt
     l0: 121
+    neuronpedia: llama3.3-70b-it/50-resid-post-gf
 
 goodfire-llama-3.1-8b-instruct:
   conversion_func: goodfire
@@ -14933,3 +14958,4 @@ goodfire-llama-3.1-8b-instruct:
   - id: layer_19
     path: Llama-3.1-8B-Instruct-SAE-l19.pth
     l0: 91
+    neuronpedia: llama3.1-8b-it/19-resid-post-gf

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/saes/gated_sae.py

@@ -2,7 +2,6 @@ from dataclasses import dataclass
 from typing import Any
 
 import torch
-from jaxtyping import Float
 from numpy.typing import NDArray
 from torch import nn
 from typing_extensions import override
@@ -49,9 +48,7 @@ class GatedSAE(SAE[GatedSAEConfig]):
         super().initialize_weights()
         _init_weights_gated(self)
 
-    def encode(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_sae"]:
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         """
         Encode the input tensor into the feature space using a gated encoder.
         This must match the original encode_gated implementation from SAE class.
@@ -72,9 +69,7 @@ class GatedSAE(SAE[GatedSAEConfig]):
         # Combine gating and magnitudes
         return self.hook_sae_acts_post(active_features * feature_magnitudes)
 
-    def decode(
-        self, feature_acts: Float[torch.Tensor, "... d_sae"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def decode(self, feature_acts: torch.Tensor) -> torch.Tensor:
         """
         Decode the feature activations back into the input space:
           1) Apply optional finetuning scaling.
@@ -147,8 +142,8 @@ class GatedTrainingSAE(TrainingSAE[GatedTrainingSAEConfig]):
         _init_weights_gated(self)
 
     def encode_with_hidden_pre(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> tuple[Float[torch.Tensor, "... d_sae"], Float[torch.Tensor, "... d_sae"]]:
+        self, x: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Gated forward pass with pre-activation (for training).
         """

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/saes/jumprelu_sae.py

@@ -3,7 +3,6 @@ from typing import Any, Literal
 
 import numpy as np
 import torch
-from jaxtyping import Float
 from torch import nn
 from typing_extensions import override
 
@@ -130,9 +129,7 @@ class JumpReLUSAE(SAE[JumpReLUSAEConfig]):
             torch.zeros(self.cfg.d_sae, dtype=self.dtype, device=self.device)
         )
 
-    def encode(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_sae"]:
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         """
         Encode the input tensor into the feature space using JumpReLU.
         The threshold parameter determines which units remain active.
@@ -150,9 +147,7 @@ class JumpReLUSAE(SAE[JumpReLUSAEConfig]):
         # 3) Multiply the normally activated units by that mask.
         return self.hook_sae_acts_post(base_acts * jump_relu_mask)
 
-    def decode(
-        self, feature_acts: Float[torch.Tensor, "... d_sae"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def decode(self, feature_acts: torch.Tensor) -> torch.Tensor:
         """
         Decode the feature activations back to the input space.
         Follows the same steps as StandardSAE: apply scaling, transform, hook, and optionally reshape.
@@ -265,8 +260,8 @@ class JumpReLUTrainingSAE(TrainingSAE[JumpReLUTrainingSAEConfig]):
         return torch.exp(self.log_threshold)
 
     def encode_with_hidden_pre(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> tuple[Float[torch.Tensor, "... d_sae"], Float[torch.Tensor, "... d_sae"]]:
+        self, x: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         sae_in = self.process_sae_in(x)
 
         hidden_pre = sae_in @ self.W_enc + self.b_enc

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/saes/matryoshka_batchtopk_sae.py

@@ -2,7 +2,6 @@ import warnings
 from dataclasses import dataclass, field
 
 import torch
-from jaxtyping import Float
 from typing_extensions import override
 
 from sae_lens.saes.batchtopk_sae import (
@@ -95,10 +94,10 @@ class MatryoshkaBatchTopKTrainingSAE(BatchTopKTrainingSAE):
 
     def _decode_matryoshka_level(
         self,
-        feature_acts: Float[torch.Tensor, "... d_sae"],
+        feature_acts: torch.Tensor,
         width: int,
         inv_W_dec_norm: torch.Tensor,
-    ) -> Float[torch.Tensor, "... d_in"]:
+    ) -> torch.Tensor:
         """
         Decodes feature activations back into input space for a matryoshka level
         """

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/saes/sae.py

@@ -19,7 +19,6 @@ from typing import (
 
 import einops
 import torch
-from jaxtyping import Float
 from numpy.typing import NDArray
 from safetensors.torch import load_file, save_file
 from torch import nn
@@ -351,16 +350,12 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         self.W_enc = nn.Parameter(w_enc_data)
 
     @abstractmethod
-    def encode(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_sae"]:
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         """Encode input tensor to feature space."""
         pass
 
     @abstractmethod
-    def decode(
-        self, feature_acts: Float[torch.Tensor, "... d_sae"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def decode(self, feature_acts: torch.Tensor) -> torch.Tensor:
         """Decode feature activations back to input space."""
         pass
 
@@ -450,9 +445,7 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
 
         return super().to(*args, **kwargs)
 
-    def process_sae_in(
-        self, sae_in: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def process_sae_in(self, sae_in: torch.Tensor) -> torch.Tensor:
         sae_in = sae_in.to(self.dtype)
         sae_in = self.reshape_fn_in(sae_in)
 
@@ -859,14 +852,12 @@ class TrainingSAE(SAE[T_TRAINING_SAE_CONFIG], ABC):
 
     @abstractmethod
     def encode_with_hidden_pre(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> tuple[Float[torch.Tensor, "... d_sae"], Float[torch.Tensor, "... d_sae"]]:
+        self, x: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         """Encode with access to pre-activation values for training."""
         ...
 
-    def encode(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_sae"]:
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         """
         For inference, just encode without returning hidden_pre.
         (training_forward_pass calls encode_with_hidden_pre).
@@ -874,9 +865,7 @@ class TrainingSAE(SAE[T_TRAINING_SAE_CONFIG], ABC):
         feature_acts, _ = self.encode_with_hidden_pre(x)
         return feature_acts
 
-    def decode(
-        self, feature_acts: Float[torch.Tensor, "... d_sae"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def decode(self, feature_acts: torch.Tensor) -> torch.Tensor:
         """
         Decodes feature activations back into input space,
         applying optional finetuning scale, hooking, out normalization, etc.

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/saes/standard_sae.py

@@ -2,7 +2,6 @@ from dataclasses import dataclass
 
 import numpy as np
 import torch
-from jaxtyping import Float
 from numpy.typing import NDArray
 from torch import nn
 from typing_extensions import override
@@ -54,9 +53,7 @@ class StandardSAE(SAE[StandardSAEConfig]):
         super().initialize_weights()
         _init_weights_standard(self)
 
-    def encode(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_sae"]:
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         """
         Encode the input tensor into the feature space.
         """
@@ -67,9 +64,7 @@ class StandardSAE(SAE[StandardSAEConfig]):
         # Apply the activation function (e.g., ReLU, depending on config)
         return self.hook_sae_acts_post(self.activation_fn(hidden_pre))
 
-    def decode(
-        self, feature_acts: Float[torch.Tensor, "... d_sae"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def decode(self, feature_acts: torch.Tensor) -> torch.Tensor:
         """
         Decode the feature activations back to the input space.
         Now, if hook_z reshaping is turned on, we reverse the flattening.
@@ -127,8 +122,8 @@ class StandardTrainingSAE(TrainingSAE[StandardTrainingSAEConfig]):
         }
 
     def encode_with_hidden_pre(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> tuple[Float[torch.Tensor, "... d_sae"], Float[torch.Tensor, "... d_sae"]]:
+        self, x: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         # Process the input (including dtype conversion, hook call, and any activation normalization)
         sae_in = self.process_sae_in(x)
         # Compute the pre-activation (and allow for a hook if desired)

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/saes/temporal_sae.py

@@ -13,7 +13,6 @@ from typing import Literal
 
 import torch
 import torch.nn.functional as F
-from jaxtyping import Float
 from torch import nn
 from typing_extensions import override
 
@@ -250,8 +249,8 @@ class TemporalSAE(SAE[TemporalSAEConfig]):
             )
 
     def encode_with_predictions(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> tuple[Float[torch.Tensor, "... d_sae"], Float[torch.Tensor, "... d_sae"]]:
+        self, x: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         """Encode input to novel codes only.
 
         Returns only the sparse novel codes (not predicted codes).
@@ -312,14 +311,10 @@ class TemporalSAE(SAE[TemporalSAEConfig]):
         # Return only novel codes (these are the interpretable features)
         return z_novel, z_pred
 
-    def encode(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_sae"]:
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         return self.encode_with_predictions(x)[0]
 
-    def decode(
-        self, feature_acts: Float[torch.Tensor, "... d_sae"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def decode(self, feature_acts: torch.Tensor) -> torch.Tensor:
         """Decode novel codes to reconstruction.
 
         Note: This only decodes the novel codes. For full reconstruction,
@@ -342,9 +337,7 @@ class TemporalSAE(SAE[TemporalSAEConfig]):
         return sae_out
 
     @override
-    def forward(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_in"]:
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
         """Full forward pass through TemporalSAE.
 
         Returns complete reconstruction (predicted + novel).

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/saes/topk_sae.py

@@ -4,7 +4,6 @@ from dataclasses import dataclass
 from typing import Any, Callable
 
 import torch
-from jaxtyping import Float
 from torch import nn
 from transformer_lens.hook_points import HookPoint
 from typing_extensions import override
@@ -235,9 +234,7 @@ class TopKSAE(SAE[TopKSAEConfig]):
         super().initialize_weights()
         _init_weights_topk(self)
 
-    def encode(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> Float[torch.Tensor, "... d_sae"]:
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         """
         Converts input x into feature activations.
         Uses topk activation under the hood.
@@ -251,8 +248,8 @@ class TopKSAE(SAE[TopKSAEConfig]):
 
     def decode(
         self,
-        feature_acts: Float[torch.Tensor, "... d_sae"],
-    ) -> Float[torch.Tensor, "... d_in"]:
+        feature_acts: torch.Tensor,
+    ) -> torch.Tensor:
         """
         Reconstructs the input from topk feature activations.
         Applies optional finetuning scaling, hooking to recons, out normalization,
@@ -354,8 +351,8 @@ class TopKTrainingSAE(TrainingSAE[TopKTrainingSAEConfig]):
         _init_weights_topk(self)
 
     def encode_with_hidden_pre(
-        self, x: Float[torch.Tensor, "... d_in"]
-    ) -> tuple[Float[torch.Tensor, "... d_sae"], Float[torch.Tensor, "... d_sae"]]:
+        self, x: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Similar to the base training method: calculate pre-activations, then apply TopK.
         """
@@ -372,8 +369,8 @@ class TopKTrainingSAE(TrainingSAE[TopKTrainingSAEConfig]):
     @override
     def decode(
         self,
-        feature_acts: Float[torch.Tensor, "... d_sae"],
-    ) -> Float[torch.Tensor, "... d_in"]:
+        feature_acts: torch.Tensor,
+    ) -> torch.Tensor:
         """
         Decodes feature activations back into input space,
         applying optional finetuning scale, hooking, out normalization, etc.

{sae_lens-6.20.1 → sae_lens-6.22.2}/sae_lens/training/activations_store.py

@@ -12,7 +12,6 @@ import torch
 from datasets import Dataset, DatasetDict, IterableDataset, load_dataset
 from huggingface_hub import hf_hub_download
 from huggingface_hub.utils import HfHubHTTPError
-from jaxtyping import Float, Int
 from requests import HTTPError
 from safetensors.torch import load_file, save_file
 from tqdm.auto import tqdm
@@ -167,9 +166,11 @@ class ActivationsStore:
         disable_concat_sequences: bool = False,
         sequence_separator_token: int | Literal["bos", "eos", "sep"] | None = "bos",
     ) -> ActivationsStore:
+        if context_size is None:
+            context_size = sae.cfg.metadata.context_size
         if sae.cfg.metadata.hook_name is None:
             raise ValueError("hook_name is required")
-        if sae.cfg.metadata.context_size is None:
+        if context_size is None:
             raise ValueError("context_size is required")
         if sae.cfg.metadata.prepend_bos is None:
             raise ValueError("prepend_bos is required")
@@ -179,9 +180,7 @@ class ActivationsStore:
             d_in=sae.cfg.d_in,
             hook_name=sae.cfg.metadata.hook_name,
             hook_head_index=sae.cfg.metadata.hook_head_index,
-            context_size=sae.cfg.metadata.context_size
-            if context_size is None
-            else context_size,
+            context_size=context_size,
             prepend_bos=sae.cfg.metadata.prepend_bos,
             streaming=streaming,
             store_batch_size_prompts=store_batch_size_prompts,
@@ -542,8 +541,8 @@ class ActivationsStore:
         d_in: int,
         raise_on_epoch_end: bool,
     ) -> tuple[
-        Float[torch.Tensor, "(total_size context_size) num_layers d_in"],
-        Int[torch.Tensor, "(total_size context_size)"] | None,
+        torch.Tensor,
+        torch.Tensor | None,
     ]:
         """
         Loads `total_size` activations from `cached_activation_dataset`