sae-lens 6.26.1__py3-none-any.whl → 6.28.2__py3-none-any.whl

sae_lens/training/mixing_buffer.py

@@ -8,15 +8,19 @@ def mixing_buffer(
     buffer_size: int,
     batch_size: int,
     activations_loader: Iterator[torch.Tensor],
+    mix_fraction: float = 0.5,
 ) -> Iterator[torch.Tensor]:
     """
     A generator that maintains a mix of old and new activations for better training.
-    It stores half of the activations and mixes them with new ones to create batches.
+    It keeps a portion of activations and mixes them with new ones to create batches.
 
     Args:
-        buffer_size: Total size of the buffer (will store buffer_size/2 activations)
+        buffer_size: Total size of the buffer
         batch_size: Size of batches to return
         activations_loader: Iterator providing new activations
+        mix_fraction: Fraction of buffer to keep for mixing (default 0.5).
+                      Higher values mean more temporal mixing but slower throughput.
+                      If 0, no shuffling occurs (passthrough mode).
 
     Yields:
         Batches of activations of shape (batch_size, *activation_dims)
@@ -24,6 +28,8 @@ def mixing_buffer(
 
     if buffer_size < batch_size:
         raise ValueError("Buffer size must be greater than or equal to batch size")
+    if not 0 <= mix_fraction <= 1:
+        raise ValueError("mix_fraction must be in [0, 1]")
 
     storage_buffer: torch.Tensor | None = None
 
@@ -35,10 +41,13 @@ def mixing_buffer(
         )
 
         if storage_buffer.shape[0] >= buffer_size:
-            # Shuffle
-            storage_buffer = storage_buffer[torch.randperm(storage_buffer.shape[0])]
+            if mix_fraction > 0:
+                storage_buffer = storage_buffer[torch.randperm(storage_buffer.shape[0])]
 
-            num_serving_batches = max(1, storage_buffer.shape[0] // (2 * batch_size))
+            # Keep a fixed amount for mixing, serve the rest
+            keep_for_mixing = int(buffer_size * mix_fraction)
+            num_to_serve = storage_buffer.shape[0] - keep_for_mixing
+            num_serving_batches = max(1, num_to_serve // batch_size)
             serving_cutoff = num_serving_batches * batch_size
             serving_buffer = storage_buffer[:serving_cutoff]
             storage_buffer = storage_buffer[serving_cutoff:]

sae_lens/training/sae_trainer.py

@@ -55,7 +55,7 @@ Evaluator = Callable[[T_TRAINING_SAE, DataProvider, ActivationScaler], dict[str,
 
 class SAETrainer(Generic[T_TRAINING_SAE, T_TRAINING_SAE_CONFIG]):
     """
-    Core SAE class used for inference. For training, see TrainingSAE.
+    Trainer for Sparse Autoencoder (SAE) models.
     """
 
     data_provider: DataProvider

sae_lens/util.py

@@ -95,8 +95,10 @@ def get_special_token_ids(tokenizer: PreTrainedTokenizerBase) -> list[int]:
     return list(special_tokens)
 
 
-def str_to_dtype(dtype: str) -> torch.dtype:
+def str_to_dtype(dtype: str | torch.dtype) -> torch.dtype:
     """Convert a string to a torch.dtype."""
+    if isinstance(dtype, torch.dtype):
+        return dtype
     if dtype not in DTYPE_MAP:
         raise ValueError(
             f"Invalid dtype: {dtype}. Must be one of {list(DTYPE_MAP.keys())}"
@@ -111,3 +113,26 @@ def dtype_to_str(dtype: torch.dtype) -> str:
             f"Invalid dtype: {dtype}. Must be one of {list(DTYPE_TO_STR.keys())}"
         )
     return DTYPE_TO_STR[dtype]
+
+
+def cosine_similarities(
+    mat1: torch.Tensor, mat2: torch.Tensor | None = None
+) -> torch.Tensor:
+    """
+    Compute cosine similarities between each row of mat1 and each row of mat2.
+
+    Args:
+        mat1: Tensor of shape [n1, d]
+        mat2: Tensor of shape [n2, d]. If not provided, mat1 = mat2
+
+    Returns:
+        Tensor of shape [n1, n2] with cosine similarities
+    """
+    if mat2 is None:
+        mat2 = mat1
+    # Clamp norm to 1e-8 to prevent division by zero. This threshold is chosen
+    # to be small enough to not affect normal vectors but large enough to avoid
+    # numerical instability. Zero vectors will effectively map to zero similarity.
+    mat1_normed = mat1 / mat1.norm(dim=1, keepdim=True).clamp(min=1e-8)
+    mat2_normed = mat2 / mat2.norm(dim=1, keepdim=True).clamp(min=1e-8)
+    return mat1_normed @ mat2_normed.T

{sae_lens-6.26.1.dist-info → sae_lens-6.28.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sae-lens
-Version: 6.26.1
+Version: 6.28.2
 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.
@@ -77,11 +79,21 @@ The new v6 update is a major refactor to SAELens and changes the way training co
   [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://githubtocolab.com/decoderesearch/SAELens/blob/main/tutorials/logits_lens_with_features.ipynb)
 - [Training a Sparse Autoencoder](tutorials/training_a_sparse_autoencoder.ipynb)
   [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://githubtocolab.com/decoderesearch/SAELens/blob/main/tutorials/training_a_sparse_autoencoder.ipynb)
+- [Training SAEs on Synthetic Data](tutorials/training_saes_on_synthetic_data.ipynb)
+  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://githubtocolab.com/decoderesearch/SAELens/blob/main/tutorials/training_saes_on_synthetic_data.ipynb)
 
 ## Join the Slack!
 
 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-6.28.2.dist-info/RECORD

@@ -0,0 +1,52 @@
+sae_lens/__init__.py,sha256=B9tY0Jt21pOHmSQrQLpMxQHyUAdLHIZpVP6pg3O0dfQ,4788
+sae_lens/analysis/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+sae_lens/analysis/hooked_sae_transformer.py,sha256=dQRgGVwce8XwylL2AzJE7l9elhtMRFCs2hdUj-Qyy4g,14038
+sae_lens/analysis/neuronpedia_integration.py,sha256=Gx1W7hUBEuMoasNcnOnZ1wmqbXDd1pSZ1nqKEya1HQc,4962
+sae_lens/cache_activations_runner.py,sha256=TjqNWIc46Nw09jHWFjzQzgzG5wdu_87Ahe-iFjI5_0Q,13117
+sae_lens/config.py,sha256=sseYcRMsAyopj8FICup1RGTXjFxzAithZ2OH7OpQV3Y,30839
+sae_lens/constants.py,sha256=CM-h9AjZNAl2aP7hVpKk7YsFHpu-_Lfhhmq2d5qPEVc,887
+sae_lens/evals.py,sha256=P0NUsJeGzYxFBiVKhbPzd72IFKY4gH40HHlEZ3jEAmg,39598
+sae_lens/llm_sae_training_runner.py,sha256=M7BK55gSFYu2qFQKABHX3c8i46P1LfODCeyHFzGGuqU,15196
+sae_lens/load_model.py,sha256=C8AMykctj6H7tz_xRwB06-EXj6TfW64PtSJZR5Jxn1Y,8649
+sae_lens/loading/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+sae_lens/loading/pretrained_sae_loaders.py,sha256=hHMlew1u6zVlbzvS9S_SfUPnAG0_OAjjIcjoUTIUZrU,63657
+sae_lens/loading/pretrained_saes_directory.py,sha256=1at_aQbD8WFywchQCKuwfP-yvCq_Z2aUYrpKDnSN5Nc,4283
+sae_lens/pretokenize_runner.py,sha256=amJwIz3CKi2s2wNQn-10E7eAV7VFhNqtFDNTeTkwEI8,7133
+sae_lens/pretrained_saes.yaml,sha256=Nq43dTcFvDDONTuJ9Me_HQ5nHqr9BdbP5-ZJGXj0TAQ,1509932
+sae_lens/registry.py,sha256=nhy7BPSudSATqW4lo9H_k3Na7sfGHmAf9v-3wpnLL_o,1490
+sae_lens/saes/__init__.py,sha256=SBqPaP6Gl5uPFwHlumAZATC4Wd26xKIYLAAAo4MSa5Q,2200
+sae_lens/saes/batchtopk_sae.py,sha256=x4EbgZl0GUickRPcCmtKNGS2Ra3Uy1Z1OtF2FnrSabQ,5422
+sae_lens/saes/gated_sae.py,sha256=V_2ZNlV4gRD-rX5JSx1xqY7idT8ChfdQ5yxWDdu_6hg,8826
+sae_lens/saes/jumprelu_sae.py,sha256=miiF-xI_yXdV9EkKjwAbU9zSMsx9KtKCz5YdXEzkN8g,13313
+sae_lens/saes/matching_pursuit_sae.py,sha256=08_G9p1YMLnE5qZVCPp6gll-iG6nHRbMMASf4_bkFt8,13207
+sae_lens/saes/matryoshka_batchtopk_sae.py,sha256=Qr6htt1HHOuO9FXI9hyaPSnGFIiJG-v7y1t1CEmkFzM,5995
+sae_lens/saes/sae.py,sha256=xRmgiLuaFlDCv8SyLbL-5TwdrWHpNLqSGe8mC1L6WcI,40942
+sae_lens/saes/standard_sae.py,sha256=_hldNZkFPAf9VGrxouR1-tN8T2OEk8IkWBcXoatrC1o,5749
+sae_lens/saes/temporal_sae.py,sha256=83Ap4mYGfdN3sKdPF8nKjhdXph3-7E2QuLobqJ_YuoM,13273
+sae_lens/saes/topk_sae.py,sha256=vrMRPrCQR1o8G_kXqY_EAoGZARupkQNFB2dNZVLsusE,21073
+sae_lens/saes/transcoder.py,sha256=CTpJs8ASOK06npih7gZHygZuxqTR7HICWlOYfTiKjI4,13501
+sae_lens/synthetic/__init__.py,sha256=FGUasB6fLPXRFCcrtKfL7vCKDOWebZ5Rx5F9QNJZklI,2875
+sae_lens/synthetic/activation_generator.py,sha256=JEN7mEgdGDuXr0ArTwUsSdSVUAfvheT_1Eew2ojbA-g,7659
+sae_lens/synthetic/correlation.py,sha256=odr-S5h6c2U-bepwrAQeMfV1iBF_cnnQzqw7zapEXZ4,6056
+sae_lens/synthetic/evals.py,sha256=Nhi314ZnRgLfhBj-3tm_zzI-pGyFTcwllDXbIpPFXeU,4584
+sae_lens/synthetic/feature_dictionary.py,sha256=ysn0ihE3JgVlCLUZMb127WYZqbz4kMp9BGHfCZqERBg,6487
+sae_lens/synthetic/firing_probabilities.py,sha256=yclz1pWl5gE1r8LAxFvzQS88Lxwk5-3r8BCX9HLVejA,3370
+sae_lens/synthetic/hierarchy.py,sha256=j9-6K7xq6zQS9N8bB5nK_-EbuzAZsY5Z5AfUK-qlB5M,22138
+sae_lens/synthetic/initialization.py,sha256=orMGW-786wRDHIS2W7bEH0HmlVFQ4g2z4bnnwdv5w4s,1386
+sae_lens/synthetic/plotting.py,sha256=5lFrej1QOkGAcImFNo5-o-8mI_rUVqvEI57KzUQPPtQ,8208
+sae_lens/synthetic/training.py,sha256=Bg6NYxdzifq_8g-dJQSZ_z_TXDdGRtEi7tqNDb-gCVc,4986
+sae_lens/tokenization_and_batching.py,sha256=uoHtAs9z3XqG0Fh-iQVYVlrbyB_E3kFFhrKU30BosCo,5438
+sae_lens/training/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+sae_lens/training/activation_scaler.py,sha256=FzNfgBplLWmyiSlZ6TUvE-nur3lOiGTrlvC97ys8S24,1973
+sae_lens/training/activations_store.py,sha256=kp4-6R4rTJUSt-g-Ifg5B1h7iIe7jZj-XQSKDvDpQMI,32187
+sae_lens/training/mixing_buffer.py,sha256=1Z-S2CcQXMWGxRZJFnXeZFxbZcALkO_fP6VO37XdJQQ,2519
+sae_lens/training/optim.py,sha256=bJpqqcK4enkcPvQAJkeH4Ci1LUOlfjIMTv6-IlaAbRA,5588
+sae_lens/training/sae_trainer.py,sha256=iiGrNwmiX0xSHnJit0lH66yQzB6q8Fww1WNJZbTSBGY,17579
+sae_lens/training/types.py,sha256=1FpLx_Doda9vZpmfm-x1e8wGBYpyhe9Kpb_JuM5nIFM,90
+sae_lens/training/upload_saes_to_huggingface.py,sha256=r_WzI1zLtGZ5TzAxuG3xa_8T09j3zXJrWd_vzPsPGkQ,4469
+sae_lens/tutorial/tsea.py,sha256=fd1am_XXsf2KMbByDapJo-2qlxduKaa62Z2qcQZ3QKU,18145
+sae_lens/util.py,sha256=oIMoeyEP2IzcPFmRbKUzOAycgEyMcOasGeO_BGVZbc4,4846
+sae_lens-6.28.2.dist-info/METADATA,sha256=i_kbAa64It0NRDrnSlmwNa8qgqOyEMntT_Ifxdx4Q90,6573
+sae_lens-6.28.2.dist-info/WHEEL,sha256=zp0Cn7JsFoX2ATtOhtaFYIiE2rmFAD4OcMhtUki8W3U,88
+sae_lens-6.28.2.dist-info/licenses/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
+sae_lens-6.28.2.dist-info/RECORD,,

sae_lens-6.26.1.dist-info/RECORD

@@ -1,42 +0,0 @@
-sae_lens/__init__.py,sha256=zRp1nmb41W1Pt1rvlKvRWw73UxjGyz1iHAzH9_X6_WQ,4725
-sae_lens/analysis/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-sae_lens/analysis/hooked_sae_transformer.py,sha256=dQRgGVwce8XwylL2AzJE7l9elhtMRFCs2hdUj-Qyy4g,14038
-sae_lens/analysis/neuronpedia_integration.py,sha256=Gx1W7hUBEuMoasNcnOnZ1wmqbXDd1pSZ1nqKEya1HQc,4962
-sae_lens/cache_activations_runner.py,sha256=Lvlz-k5-3XxVRtUdC4b1CiKyx5s0ckLa8GDGv9_kcxs,12566
-sae_lens/config.py,sha256=C982bUELhGHcfTwzeMTtXIf2hPtc946thYpUyctLiBo,30516
-sae_lens/constants.py,sha256=CM-h9AjZNAl2aP7hVpKk7YsFHpu-_Lfhhmq2d5qPEVc,887
-sae_lens/evals.py,sha256=P0NUsJeGzYxFBiVKhbPzd72IFKY4gH40HHlEZ3jEAmg,39598
-sae_lens/llm_sae_training_runner.py,sha256=M7BK55gSFYu2qFQKABHX3c8i46P1LfODCeyHFzGGuqU,15196
-sae_lens/load_model.py,sha256=C8AMykctj6H7tz_xRwB06-EXj6TfW64PtSJZR5Jxn1Y,8649
-sae_lens/loading/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-sae_lens/loading/pretrained_sae_loaders.py,sha256=hq-dhxsEdUmlAnZEiZBqX7lNyQQwZ6KXmXZWpzAc5FY,63638
-sae_lens/loading/pretrained_saes_directory.py,sha256=hejNfLUepYCSGPalRfQwxxCEUqMMUPsn1tufwvwct5k,3820
-sae_lens/pretokenize_runner.py,sha256=amJwIz3CKi2s2wNQn-10E7eAV7VFhNqtFDNTeTkwEI8,7133
-sae_lens/pretrained_saes.yaml,sha256=Hn8jXwZ7V6QQxzgu41LFEP-LAzuDxwYL5vhoar-pPX8,1509922
-sae_lens/registry.py,sha256=nhy7BPSudSATqW4lo9H_k3Na7sfGHmAf9v-3wpnLL_o,1490
-sae_lens/saes/__init__.py,sha256=SBqPaP6Gl5uPFwHlumAZATC4Wd26xKIYLAAAo4MSa5Q,2200
-sae_lens/saes/batchtopk_sae.py,sha256=x4EbgZl0GUickRPcCmtKNGS2Ra3Uy1Z1OtF2FnrSabQ,5422
-sae_lens/saes/gated_sae.py,sha256=mHnmw-RD7hqIbP9_EBj3p2SK0OqQIkZivdOKRygeRgw,8825
-sae_lens/saes/jumprelu_sae.py,sha256=udjGHp3WTABQSL2Qq57j-bINWX61GCmo68EmdjMOXoo,13310
-sae_lens/saes/matching_pursuit_sae.py,sha256=08_G9p1YMLnE5qZVCPp6gll-iG6nHRbMMASf4_bkFt8,13207
-sae_lens/saes/matryoshka_batchtopk_sae.py,sha256=Qr6htt1HHOuO9FXI9hyaPSnGFIiJG-v7y1t1CEmkFzM,5995
-sae_lens/saes/sae.py,sha256=fzXv8lwHskSxsf8hm_wlKPkpq50iafmBjBNQzwZ6a00,40050
-sae_lens/saes/standard_sae.py,sha256=nEVETwAmRD2tyX7ESIic1fij48gAq1Dh7s_GQ2fqCZ4,5747
-sae_lens/saes/temporal_sae.py,sha256=DsecivcHWId-MTuJpQbz8OhqtmGhZACxJauYZGHo0Ok,13272
-sae_lens/saes/topk_sae.py,sha256=vrMRPrCQR1o8G_kXqY_EAoGZARupkQNFB2dNZVLsusE,21073
-sae_lens/saes/transcoder.py,sha256=CTpJs8ASOK06npih7gZHygZuxqTR7HICWlOYfTiKjI4,13501
-sae_lens/tokenization_and_batching.py,sha256=D_o7cXvRqhT89H3wNzoRymNALNE6eHojBWLdXOUwUGE,5438
-sae_lens/training/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-sae_lens/training/activation_scaler.py,sha256=FzNfgBplLWmyiSlZ6TUvE-nur3lOiGTrlvC97ys8S24,1973
-sae_lens/training/activations_store.py,sha256=rQadexm2BiwK7_MZIPlRkcKSqabi3iuOTC-R8aJchS8,33778
-sae_lens/training/mixing_buffer.py,sha256=vDpYG5ZE70szDvBsRKcNHEES3h_WTKJ16qDYk5jPOVA,2015
-sae_lens/training/optim.py,sha256=bJpqqcK4enkcPvQAJkeH4Ci1LUOlfjIMTv6-IlaAbRA,5588
-sae_lens/training/sae_trainer.py,sha256=zhkabyIKxI_tZTV3_kwz6zMrHZ95Ecr97krmwc-9ffs,17600
-sae_lens/training/types.py,sha256=1FpLx_Doda9vZpmfm-x1e8wGBYpyhe9Kpb_JuM5nIFM,90
-sae_lens/training/upload_saes_to_huggingface.py,sha256=r_WzI1zLtGZ5TzAxuG3xa_8T09j3zXJrWd_vzPsPGkQ,4469
-sae_lens/tutorial/tsea.py,sha256=fd1am_XXsf2KMbByDapJo-2qlxduKaa62Z2qcQZ3QKU,18145
-sae_lens/util.py,sha256=spkcmQUsjVYFn5H2032nQYr1CKGVnv3tAdfIpY59-Mg,3919
-sae_lens-6.26.1.dist-info/METADATA,sha256=yoE6CFgQ9L5SLzI3Zgr8H8CfUBgSimihGyEIvKd8TW8,5361
-sae_lens-6.26.1.dist-info/WHEEL,sha256=zp0Cn7JsFoX2ATtOhtaFYIiE2rmFAD4OcMhtUki8W3U,88
-sae_lens-6.26.1.dist-info/licenses/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
-sae_lens-6.26.1.dist-info/RECORD,,