sae-lens 6.5.1__py3-none-any.whl → 6.5.3__py3-none-any.whl

sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.5.1"
+__version__ = "6.5.3"
 
 import logging
 

sae_lens/evals.py

@@ -776,6 +776,7 @@ def multiple_evals(
     n_eval_sparsity_variance_batches: int,
     eval_batch_size_prompts: int = 8,
     datasets: list[str] = ["Skylion007/openwebtext", "lighteval/MATH"],
+    dataset_trust_remote_code: bool = False,
     ctx_lens: list[int] = [128],
     output_dir: str = "eval_results",
     verbose: bool = False,
@@ -822,7 +823,11 @@ def multiple_evals(
         for ctx_len in ctx_lens:
             for dataset in datasets:
                 activation_store = ActivationsStore.from_sae(
-                    current_model, sae, context_size=ctx_len, dataset=dataset
+                    current_model,
+                    sae,
+                    context_size=ctx_len,
+                    dataset=dataset,
+                    dataset_trust_remote_code=dataset_trust_remote_code,
                 )
                 activation_store.shuffle_input_dataset(seed=42)
 
@@ -882,6 +887,7 @@ def run_evaluations(args: argparse.Namespace) -> list[dict[str, Any]]:
         n_eval_sparsity_variance_batches=args.n_eval_sparsity_variance_batches,
         eval_batch_size_prompts=args.batch_size_prompts,
         datasets=args.datasets,
+        dataset_trust_remote_code=args.dataset_trust_remote_code,
         ctx_lens=args.ctx_lens,
         output_dir=args.output_dir,
         verbose=args.verbose,
@@ -1004,6 +1010,11 @@ def process_args(args: list[str]) -> argparse.Namespace:
         default=["Skylion007/openwebtext"],
         help="Datasets to evaluate on, such as 'Skylion007/openwebtext' or 'lighteval/MATH'.",
     )
+    arg_parser.add_argument(
+        "--dataset_trust_remote_code",
+        action="store_true",
+        help="Allow execution of remote code when loading datasets for evaluation.",
+    )
     arg_parser.add_argument(
         "--ctx_lens",
         nargs="+",

sae_lens/load_model.py

@@ -52,7 +52,7 @@ def load_model(
     if model_class_name == "AutoModelForCausalLM":
         hf_model = AutoModelForCausalLM.from_pretrained(
             model_name, **model_from_pretrained_kwargs
-        ).to(device)
+        ).to(device)  # type: ignore
         tokenizer = AutoTokenizer.from_pretrained(model_name)
         return HookedProxyLM(hf_model, tokenizer)
 

sae_lens/loading/pretrained_sae_loaders.py

@@ -1177,6 +1177,7 @@ def get_gemma_2_transcoder_config_from_hf(
         "prepend_bos": True,
         "dataset_path": "monology/pile-uncopyrighted",
         "context_size": 1024,
+        "apply_b_dec_to_input": False,
         **(cfg_overrides or {}),
     }
 

sae_lens/saes/transcoder.py

@@ -52,6 +52,11 @@ class TranscoderConfig(SAEConfig):
 
         return res
 
+    def __post_init__(self):
+        if self.apply_b_dec_to_input:
+            raise ValueError("apply_b_dec_to_input is not supported for transcoders")
+        return super().__post_init__()
+
 
 class Transcoder(SAE[TranscoderConfig]):
     """

{sae_lens-6.5.1.dist-info → sae_lens-6.5.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: sae-lens
-Version: 6.5.1
+Version: 6.5.3
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 Keywords: deep-learning,sparse-autoencoders,mechanistic-interpretability,PyTorch
@@ -16,7 +16,7 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Provides-Extra: mamba
 Requires-Dist: automated-interpretability (>=0.0.5,<1.0.0)
 Requires-Dist: babe (>=0.0.7,<0.0.8)
-Requires-Dist: datasets (>=2.17.1,<3.0.0)
+Requires-Dist: datasets (>=3.1.0,<4.0.0)
 Requires-Dist: mamba-lens (>=0.0.4,<0.0.5) ; extra == "mamba"
 Requires-Dist: matplotlib (>=3.8.3,<4.0.0)
 Requires-Dist: matplotlib-inline (>=0.1.6,<0.2.0)

{sae_lens-6.5.1.dist-info → sae_lens-6.5.3.dist-info}/RECORD

@@ -1,15 +1,15 @@
-sae_lens/__init__.py,sha256=v3iyvnzUIs3XkzrtiW_D0FH9tDRqcxpPEzlBdGkhiRo,3588
+sae_lens/__init__.py,sha256=HBSxvQf7TM7R4n1Al0RbtS0bVFgfb3eDBMuLXZs2cKw,3588
 sae_lens/analysis/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sae_lens/analysis/hooked_sae_transformer.py,sha256=Eyg1Y2hVIHNuiiLOCTgzstOuW6iA-7hPHqaGR8y_vMs,13809
 sae_lens/analysis/neuronpedia_integration.py,sha256=MrENqc81Mc2SMbxGjbwHzpkGUCAFKSf0i4EdaUF2Oj4,18707
 sae_lens/cache_activations_runner.py,sha256=L5hhuU2-zPQr2S3L64GMKKLeMQfqXxwDl8NbuOtrybI,12567
 sae_lens/config.py,sha256=IrjbsKBbaZoFXYrsPJ5xBwIqi9uZJIIFXjV_uoErJaE,28176
 sae_lens/constants.py,sha256=CSjmiZ-bhjQeVLyRvWxAjBokCgkfM8mnvd7-vxLIWTY,639
-sae_lens/evals.py,sha256=kQyrzczKaVD9rHwfFa_DxL_gMXDxsoIVHmsFIPIU2bY,38696
+sae_lens/evals.py,sha256=rVmKqdSa4xlZw-78qpgdhDleJIO1EI1QHRiPsObW1Tc,39129
 sae_lens/llm_sae_training_runner.py,sha256=exxNX_OEhdiUrlgmBP9bjX9DOf0HUcNQGO4unKeDjKM,13713
-sae_lens/load_model.py,sha256=dBB_9gO6kWyQ4sXHq7qB8T3YUlXm3PGwYcpR4UVW4QY,8633
+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=RfZhE8l2y5_ZW5-fW-iN-SJGJfHRaZo4d6AiooK1Xuc,40890
+sae_lens/loading/pretrained_sae_loaders.py,sha256=0nb3PG26XG1EEvVIf2Jp-4_g2E30XulA5u5ok4Ah8no,40929
 sae_lens/loading/pretrained_saes_directory.py,sha256=4Vn-Jex6SveD7EbxcSOBv8cx1gkPfUMLU1QOP-ww1ZE,3752
 sae_lens/pretokenize_runner.py,sha256=w0f6SfZLAxbp5eAAKnet8RqUB_DKofZ9RGsoJwFnYbA,7058
 sae_lens/pretrained_saes.yaml,sha256=BncZpkf-NhVOWV1FhrDPr_TPxAINP4-eiPZBw-HO4N4,578197
@@ -21,7 +21,7 @@ sae_lens/saes/jumprelu_sae.py,sha256=3xkhBcCol2mEpIBLceymCpudocm2ypOjTeTXbpiXoA4
 sae_lens/saes/sae.py,sha256=gdUZuLaOHQrPjbDj-nZI813B6-_mNAnV9i9z4qTnpHk,38255
 sae_lens/saes/standard_sae.py,sha256=9UqYyYtQuThYxXKNaDjYcyowpOx2-7cShG-TeUP6JCQ,5940
 sae_lens/saes/topk_sae.py,sha256=CXMBI6CFvI5829bOhoQ350VXR9d8uFHUDlULTIWHXoU,8686
-sae_lens/saes/transcoder.py,sha256=qbsvIeRy7M1nuHapDut0URNg9fR-UtfBvwW4_hgsAac,11963
+sae_lens/saes/transcoder.py,sha256=BfLSbTYVNZh-ruGxseZiZJ_acEL6_7QyTdfqUr0lDOg,12156
 sae_lens/tokenization_and_batching.py,sha256=now7caLbU3p-iGokNwmqZDyIvxYoXgnG1uklhgiLZN4,4656
 sae_lens/training/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sae_lens/training/activation_scaler.py,sha256=seEE-2Qd2JMHxqgnsNWPt-DGtYGZxWPnOwCGuVNSOtI,1719
@@ -33,7 +33,7 @@ sae_lens/training/types.py,sha256=qSjmGzXf3MLalygG0psnVjmhX_mpLmL47MQtZfe7qxg,81
 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=mCwLAilGMVo8Scm7CIsCafU7GsfmBvCcjwmloI4Ly7Y,1718
-sae_lens-6.5.1.dist-info/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
-sae_lens-6.5.1.dist-info/METADATA,sha256=wpZeKz6pPucYSC_9HOM3FR_aq8rxPSKxn2WW4rnz7Vo,5555
-sae_lens-6.5.1.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
-sae_lens-6.5.1.dist-info/RECORD,,
+sae_lens-6.5.3.dist-info/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
+sae_lens-6.5.3.dist-info/METADATA,sha256=LP8y6l6lH8Ibsm9_HxTX-iem1uKUsailabbkBii5YIk,5554
+sae_lens-6.5.3.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+sae_lens-6.5.3.dist-info/RECORD,,