sae-lens 6.28.0__tar.gz → 6.29.0__tar.gz

{sae_lens-6.28.0 → sae_lens-6.29.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sae-lens
-Version: 6.28.0
+Version: 6.29.0
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 License-File: LICENSE
@@ -50,6 +50,8 @@ SAELens exists to help researchers:
 - Analyse sparse autoencoders / research mechanistic interpretability.
 - Generate insights which make it easier to create safe and aligned AI systems.
 
+SAELens inference works with any PyTorch-based model, not just TransformerLens. While we provide deep integration with TransformerLens via `HookedSAETransformer`, SAEs can be used with Hugging Face Transformers, NNsight, or any other framework by extracting activations and passing them to the SAE's `encode()` and `decode()` methods.
+
 Please refer to the [documentation](https://decoderesearch.github.io/SAELens/) for information on how to:
 
 - Download and Analyse pre-trained sparse autoencoders.
@@ -84,6 +86,14 @@ The new v6 update is a major refactor to SAELens and changes the way training co
 
 Feel free to join the [Open Source Mechanistic Interpretability Slack](https://join.slack.com/t/opensourcemechanistic/shared_invite/zt-375zalm04-GFd5tdBU1yLKlu_T_JSqZQ) for support!
 
+## Other SAE Projects
+
+- [dictionary-learning](https://github.com/saprmarks/dictionary_learning): An SAE training library that focuses on having hackable code.
+- [Sparsify](https://github.com/EleutherAI/sparsify): A lean SAE training library focused on TopK SAEs.
+- [Overcomplete](https://github.com/KempnerInstitute/overcomplete): SAE training library focused on vision models.
+- [SAE-Vis](https://github.com/callummcdougall/sae_vis): A library for visualizing SAE features, works with SAELens.
+- [SAEBench](https://github.com/adamkarvonen/SAEBench): A suite of LLM SAE benchmarks, works with SAELens.
+
 ## Citation
 
 Please cite the package as follows:

{sae_lens-6.28.0 → sae_lens-6.29.0}/README.md

@@ -14,6 +14,8 @@ SAELens exists to help researchers:
 - Analyse sparse autoencoders / research mechanistic interpretability.
 - Generate insights which make it easier to create safe and aligned AI systems.
 
+SAELens inference works with any PyTorch-based model, not just TransformerLens. While we provide deep integration with TransformerLens via `HookedSAETransformer`, SAEs can be used with Hugging Face Transformers, NNsight, or any other framework by extracting activations and passing them to the SAE's `encode()` and `decode()` methods.
+
 Please refer to the [documentation](https://decoderesearch.github.io/SAELens/) for information on how to:
 
 - Download and Analyse pre-trained sparse autoencoders.
@@ -48,6 +50,14 @@ The new v6 update is a major refactor to SAELens and changes the way training co
 
 Feel free to join the [Open Source Mechanistic Interpretability Slack](https://join.slack.com/t/opensourcemechanistic/shared_invite/zt-375zalm04-GFd5tdBU1yLKlu_T_JSqZQ) for support!
 
+## Other SAE Projects
+
+- [dictionary-learning](https://github.com/saprmarks/dictionary_learning): An SAE training library that focuses on having hackable code.
+- [Sparsify](https://github.com/EleutherAI/sparsify): A lean SAE training library focused on TopK SAEs.
+- [Overcomplete](https://github.com/KempnerInstitute/overcomplete): SAE training library focused on vision models.
+- [SAE-Vis](https://github.com/callummcdougall/sae_vis): A library for visualizing SAE features, works with SAELens.
+- [SAEBench](https://github.com/adamkarvonen/SAEBench): A suite of LLM SAE benchmarks, works with SAELens.
+
 ## Citation
 
 Please cite the package as follows:

{sae_lens-6.28.0 → sae_lens-6.29.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "sae-lens"
-version = "6.28.0"
+version = "6.29.0"
 description = "Training and Analyzing Sparse Autoencoders (SAEs)"
 authors = ["Joseph Bloom"]
 readme = "README.md"

{sae_lens-6.28.0 → sae_lens-6.29.0}/sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.28.0"
+__version__ = "6.29.0"
 
 import logging
 

{sae_lens-6.28.0 → sae_lens-6.29.0}/sae_lens/loading/pretrained_saes_directory.py

@@ -103,3 +103,21 @@ def get_config_overrides(release: str, sae_id: str) -> dict[str, Any]:
         if sae_info.neuronpedia_id is not None and sae_id in sae_info.neuronpedia_id:
             config_overrides["neuronpedia_id"] = sae_info.neuronpedia_id[sae_id]
     return config_overrides
+
+
+def get_releases_for_repo_id(repo_id: str) -> list[str]:
+    """
+    Find all release names that use the given HuggingFace repo_id.
+
+    Args:
+        repo_id: The HuggingFace repository ID to search for.
+
+    Returns:
+        A list of release names that use this repo_id.
+    """
+    saes_directory = get_pretrained_saes_directory()
+    return [
+        release
+        for release, lookup in saes_directory.items()
+        if lookup.repo_id == repo_id
+    ]

{sae_lens-6.28.0 → sae_lens-6.29.0}/sae_lens/pretrained_saes.yaml

@@ -40631,7 +40631,7 @@ gemma-3-1b-res-matryoshka-dc:
   conversion_func: null
   links:
     model: https://huggingface.co/google/gemma-3-1b-pt
-  model: gemma-3-1b
+  model: google/gemma-3-1b-pt
   repo_id: chanind/gemma-3-1b-batch-topk-matryoshka-saes-w-32k-l0-40
   saes:
   - id: blocks.0.hook_resid_post

{sae_lens-6.28.0 → sae_lens-6.29.0}/sae_lens/saes/sae.py

@@ -47,6 +47,7 @@ from sae_lens.loading.pretrained_saes_directory import (
     get_config_overrides,
     get_norm_scaling_factor,
     get_pretrained_saes_directory,
+    get_releases_for_repo_id,
     get_repo_id_and_folder_name,
 )
 from sae_lens.registry import get_sae_class, get_sae_training_class
@@ -624,6 +625,18 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
                 raise ValueError(
                     f"Release {release} not found in pretrained SAEs directory, and is not a valid huggingface repo."
                 )
+            # Check if the user passed a repo_id that's in the pretrained SAEs list
+            matching_releases = get_releases_for_repo_id(release)
+            if matching_releases:
+                warnings.warn(
+                    f"You are loading an SAE using the HuggingFace repo_id '{release}' directly. "
+                    f"This repo is registered in the official pretrained SAEs list with release name(s): {matching_releases}. "
+                    f"For better compatibility and to access additional metadata, consider loading with: "
+                    f"SAE.from_pretrained(release='{matching_releases[0]}', sae_id='<sae_id>'). "
+                    f"See the full list of pretrained SAEs at: https://decoderesearch.github.io/SAELens/latest/pretrained_saes/",
+                    UserWarning,
+                    stacklevel=2,
+                )
         elif sae_id not in sae_directory[release].saes_map:
             # Handle special cases like Gemma Scope
             if (

{sae_lens-6.28.0 → sae_lens-6.29.0}/sae_lens/synthetic/__init__.py

@@ -17,11 +17,14 @@ from sae_lens.synthetic.activation_generator import (
     ActivationGenerator,
     ActivationsModifier,
     ActivationsModifierInput,
+    CorrelationMatrixInput,
 )
 from sae_lens.synthetic.correlation import (
+    LowRankCorrelationMatrix,
     create_correlation_matrix_from_correlations,
     generate_random_correlation_matrix,
     generate_random_correlations,
+    generate_random_low_rank_correlation_matrix,
 )
 from sae_lens.synthetic.evals import (
     SyntheticDataEvalResult,
@@ -66,6 +69,9 @@ __all__ = [
     "create_correlation_matrix_from_correlations",
     "generate_random_correlations",
     "generate_random_correlation_matrix",
+    "generate_random_low_rank_correlation_matrix",
+    "LowRankCorrelationMatrix",
+    "CorrelationMatrixInput",
     # Feature modifiers
     "ActivationsModifier",
     "ActivationsModifierInput",

{sae_lens-6.28.0 → sae_lens-6.29.0}/sae_lens/synthetic/activation_generator.py

@@ -7,12 +7,16 @@ from collections.abc import Callable, Sequence
 import torch
 from scipy.stats import norm
 from torch import nn
-from torch.distributions import MultivariateNormal
+from torch.distributions import LowRankMultivariateNormal, MultivariateNormal
 
+from sae_lens.synthetic.correlation import LowRankCorrelationMatrix
 from sae_lens.util import str_to_dtype
 
 ActivationsModifier = Callable[[torch.Tensor], torch.Tensor]
 ActivationsModifierInput = ActivationsModifier | Sequence[ActivationsModifier] | None
+CorrelationMatrixInput = (
+    torch.Tensor | LowRankCorrelationMatrix | tuple[torch.Tensor, torch.Tensor]
+)
 
 
 class ActivationGenerator(nn.Module):
@@ -28,6 +32,7 @@ class ActivationGenerator(nn.Module):
     mean_firing_magnitudes: torch.Tensor
     modify_activations: ActivationsModifier | None
     correlation_matrix: torch.Tensor | None
+    low_rank_correlation: tuple[torch.Tensor, torch.Tensor] | None
     correlation_thresholds: torch.Tensor | None
 
     def __init__(
@@ -37,7 +42,7 @@ class ActivationGenerator(nn.Module):
         std_firing_magnitudes: torch.Tensor | float = 0.0,
         mean_firing_magnitudes: torch.Tensor | float = 1.0,
         modify_activations: ActivationsModifierInput = None,
-        correlation_matrix: torch.Tensor | None = None,
+        correlation_matrix: CorrelationMatrixInput | None = None,
         device: torch.device | str = "cpu",
         dtype: torch.dtype | str = "float32",
     ):
@@ -54,15 +59,32 @@ class ActivationGenerator(nn.Module):
         )
         self.modify_activations = _normalize_modifiers(modify_activations)
         self.correlation_thresholds = None
+        self.correlation_matrix = None
+        self.low_rank_correlation = None
+
         if correlation_matrix is not None:
-            _validate_correlation_matrix(correlation_matrix, num_features)
+            if isinstance(correlation_matrix, torch.Tensor):
+                # Full correlation matrix
+                _validate_correlation_matrix(correlation_matrix, num_features)
+                self.correlation_matrix = correlation_matrix
+            else:
+                # Low-rank correlation (tuple or LowRankCorrelationMatrix)
+                correlation_factor, correlation_diag = (
+                    correlation_matrix[0],
+                    correlation_matrix[1],
+                )
+                _validate_low_rank_correlation(
+                    correlation_factor, correlation_diag, num_features
+                )
+                self.low_rank_correlation = (correlation_factor, correlation_diag)
+
             self.correlation_thresholds = torch.tensor(
                 [norm.ppf(1 - p.item()) for p in self.firing_probabilities],
                 device=device,
                 dtype=self.firing_probabilities.dtype,
             )
-        self.correlation_matrix = correlation_matrix
 
+    @torch.no_grad()
     def sample(self, batch_size: int) -> torch.Tensor:
         """
         Generate a batch of feature activations with controlled properties.
@@ -89,6 +111,15 @@ class ActivationGenerator(nn.Module):
                 self.correlation_thresholds,
                 device,
             )
+        elif self.low_rank_correlation is not None:
+            assert self.correlation_thresholds is not None
+            firing_features = _generate_low_rank_correlated_features(
+                batch_size,
+                self.low_rank_correlation[0],
+                self.low_rank_correlation[1],
+                self.correlation_thresholds,
+                device,
+            )
         else:
             firing_features = torch.bernoulli(
                 self.firing_probabilities.unsqueeze(0).expand(batch_size, -1)
@@ -132,7 +163,7 @@ def _generate_correlated_features(
         device: Device to generate samples on
 
     Returns:
-        Binary feature matrix of shape [batch_size, num_features]
+        Binary feature matrix of shape (batch_size, num_features)
     """
     num_features = correlation_matrix.shape[0]
 
@@ -145,6 +176,41 @@ def _generate_correlated_features(
     return (gaussian_samples > thresholds.unsqueeze(0)).float()
 
 
+def _generate_low_rank_correlated_features(
+    batch_size: int,
+    correlation_factor: torch.Tensor,
+    correlation_diag: torch.Tensor,
+    thresholds: torch.Tensor,
+    device: torch.device,
+) -> torch.Tensor:
+    """
+    Generate correlated binary features using low-rank multivariate Gaussian sampling.
+
+    Uses the Gaussian copula approach with a low-rank covariance structure for scalability.
+    The covariance is represented as: cov = factor @ factor.T + diag(diag_term)
+
+    Args:
+        batch_size: Number of samples to generate
+        correlation_factor: Factor matrix of shape (num_features, rank)
+        correlation_diag: Diagonal term of shape (num_features,)
+        thresholds: Pre-computed thresholds for each feature (from inverse normal CDF)
+        device: Device to generate samples on
+
+    Returns:
+        Binary feature matrix of shape (batch_size, num_features)
+    """
+    mvn = LowRankMultivariateNormal(
+        loc=torch.zeros(
+            correlation_factor.shape[0], device=device, dtype=thresholds.dtype
+        ),
+        cov_factor=correlation_factor.to(device=device, dtype=thresholds.dtype),
+        cov_diag=correlation_diag.to(device=device, dtype=thresholds.dtype),
+    )
+
+    gaussian_samples = mvn.sample((batch_size,))
+    return (gaussian_samples > thresholds.unsqueeze(0)).float()
+
+
 def _to_tensor(
     value: torch.Tensor | float,
     num_features: int,
@@ -193,7 +259,7 @@ def _validate_correlation_matrix(
 
     Args:
         correlation_matrix: The matrix to validate
-        num_features: Expected number of features (matrix should be [num_features, num_features])
+        num_features: Expected number of features (matrix should be (num_features, num_features))
 
     Raises:
         ValueError: If the matrix has incorrect shape, non-unit diagonal, or is not positive definite
@@ -213,3 +279,36 @@ def _validate_correlation_matrix(
         torch.linalg.cholesky(correlation_matrix)
     except RuntimeError as e:
         raise ValueError("Correlation matrix must be positive definite") from e
+
+
+def _validate_low_rank_correlation(
+    correlation_factor: torch.Tensor,
+    correlation_diag: torch.Tensor,
+    num_features: int,
+) -> None:
+    """Validate that low-rank correlation parameters have correct properties.
+
+    Args:
+        correlation_factor: Factor matrix of shape (num_features, rank)
+        correlation_diag: Diagonal term of shape (num_features,)
+        num_features: Expected number of features
+
+    Raises:
+        ValueError: If shapes are incorrect or diagonal terms are not positive
+    """
+    if correlation_factor.ndim != 2:
+        raise ValueError(
+            f"correlation_factor must be 2D, got {correlation_factor.ndim}D"
+        )
+    if correlation_factor.shape[0] != num_features:
+        raise ValueError(
+            f"correlation_factor must have shape ({num_features}, rank), "
+            f"got {tuple(correlation_factor.shape)}"
+        )
+    if correlation_diag.shape != (num_features,):
+        raise ValueError(
+            f"correlation_diag must have shape ({num_features},), "
+            f"got {tuple(correlation_diag.shape)}"
+        )
+    if torch.any(correlation_diag <= 0):
+        raise ValueError("correlation_diag must have all positive values")