sae-lens 6.11.0__py3-none-any.whl → 6.12.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (7) hide show
  1. sae_lens/__init__.py +1 -1
  2. sae_lens/pretrained_saes.yaml +55 -1
  3. sae_lens/saes/batchtopk_sae.py +3 -2
  4. {sae_lens-6.11.0.dist-info → sae_lens-6.12.0.dist-info}/METADATA +1 -1
  5. {sae_lens-6.11.0.dist-info → sae_lens-6.12.0.dist-info}/RECORD +7 -7
  6. {sae_lens-6.11.0.dist-info → sae_lens-6.12.0.dist-info}/LICENSE +0 -0
  7. {sae_lens-6.11.0.dist-info → sae_lens-6.12.0.dist-info}/WHEEL +0 -0
sae_lens/__init__.py CHANGED
@@ -1,5 +1,5 @@
1
1
  # ruff: noqa: E402
2
- __version__ = "6.11.0"
2
+ __version__ = "6.12.0"
3
3
 
4
4
  import logging
5
5
 
sae_lens/pretrained_saes.yaml CHANGED
@@ -14828,4 +14828,58 @@ mntss-gemma-2-2b-2.5m-clt-as-per-layer:
14828
14828
  neuronpedia: gemma-2-2b/24-clt-hp
14829
14829
  - id: layer_25
14830
14830
  path: 25
14831
- neuronpedia: gemma-2-2b/25-clt-hp
14831
+ neuronpedia: gemma-2-2b/25-clt-hp
14832
+
14833
+ llama-3.1-8b-instruct-andyrdt:
14834
+ conversion_func: dictionary_learning_1
14835
+ model: meta-llama/Llama-3.1-8B-Instruct
14836
+ repo_id: andyrdt/saes-llama-3.1-8b-instruct
14837
+ saes:
14838
+ - id: resid_post_layer_3_trainer_1
14839
+ path: resid_post_layer_3/trainer_1
14840
+ neuronpedia: llama3.1-8b-it/3-resid-post-ar
14841
+ - id: resid_post_layer_7_trainer_1
14842
+ path: resid_post_layer_7/trainer_1
14843
+ neuronpedia: llama3.1-8b-it/7-resid-post-ar
14844
+ - id: resid_post_layer_11_trainer_1
14845
+ path: resid_post_layer_11/trainer_1
14846
+ neuronpedia: llama3.1-8b-it/11-resid-post-ar
14847
+ - id: resid_post_layer_15_trainer_1
14848
+ path: resid_post_layer_15/trainer_1
14849
+ neuronpedia: llama3.1-8b-it/15-resid-post-ar
14850
+ - id: resid_post_layer_19_trainer_1
14851
+ path: resid_post_layer_19/trainer_1
14852
+ neuronpedia: llama3.1-8b-it/19-resid-post-ar
14853
+ - id: resid_post_layer_23_trainer_1
14854
+ path: resid_post_layer_23/trainer_1
14855
+ neuronpedia: llama3.1-8b-it/23-resid-post-ar
14856
+ - id: resid_post_layer_27_trainer_1
14857
+ path: resid_post_layer_27/trainer_1
14858
+ neuronpedia: llama3.1-8b-it/27-resid-post-ar
14859
+
14860
+ qwen2.5-7b-instruct-andyrdt:
14861
+ conversion_func: dictionary_learning_1
14862
+ model: Qwen/Qwen2.5-7B-Instruct
14863
+ repo_id: andyrdt/saes-qwen2.5-7b-instruct
14864
+ saes:
14865
+ - id: resid_post_layer_3_trainer_1
14866
+ path: resid_post_layer_3/trainer_1
14867
+ neuronpedia: qwen2.5-7b-it/3-resid-post-ar
14868
+ - id: resid_post_layer_7_trainer_1
14869
+ path: resid_post_layer_7/trainer_1
14870
+ neuronpedia: qwen2.5-7b-it/7-resid-post-ar
14871
+ - id: resid_post_layer_11_trainer_1
14872
+ path: resid_post_layer_11/trainer_1
14873
+ neuronpedia: qwen2.5-7b-it/11-resid-post-ar
14874
+ - id: resid_post_layer_15_trainer_1
14875
+ path: resid_post_layer_15/trainer_1
14876
+ neuronpedia: qwen2.5-7b-it/15-resid-post-ar
14877
+ - id: resid_post_layer_19_trainer_1
14878
+ path: resid_post_layer_19/trainer_1
14879
+ neuronpedia: qwen2.5-7b-it/19-resid-post-ar
14880
+ - id: resid_post_layer_23_trainer_1
14881
+ path: resid_post_layer_23/trainer_1
14882
+ neuronpedia: qwen2.5-7b-it/23-resid-post-ar
14883
+ - id: resid_post_layer_27_trainer_1
14884
+ path: resid_post_layer_27/trainer_1
14885
+ neuronpedia: qwen2.5-7b-it/27-resid-post-ar
sae_lens/saes/batchtopk_sae.py CHANGED
@@ -15,7 +15,7 @@ class BatchTopK(nn.Module):
15
15
 
16
16
  def __init__(
17
17
  self,
18
- k: int,
18
+ k: float,
19
19
  ):
20
20
  super().__init__()
21
21
  self.k = k
@@ -23,7 +23,7 @@ class BatchTopK(nn.Module):
23
23
  def forward(self, x: torch.Tensor) -> torch.Tensor:
24
24
  acts = x.relu()
25
25
  flat_acts = acts.flatten()
26
- acts_topk_flat = torch.topk(flat_acts, self.k * acts.shape[0], dim=-1)
26
+ acts_topk_flat = torch.topk(flat_acts, int(self.k * acts.shape[0]), dim=-1)
27
27
  return (
28
28
  torch.zeros_like(flat_acts)
29
29
  .scatter(-1, acts_topk_flat.indices, acts_topk_flat.values)
@@ -37,6 +37,7 @@ class BatchTopKTrainingSAEConfig(TopKTrainingSAEConfig):
37
37
  Configuration class for training a BatchTopKTrainingSAE.
38
38
  """
39
39
 
40
+ k: float = 100 # type: ignore[assignment]
40
41
  topk_threshold_lr: float = 0.01
41
42
 
42
43
  @override
{sae_lens-6.11.0.dist-info → sae_lens-6.12.0.dist-info}/METADATA RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.3
2
2
  Name: sae-lens
3
- Version: 6.11.0
3
+ Version: 6.12.0
4
4
  Summary: Training and Analyzing Sparse Autoencoders (SAEs)
5
5
  License: MIT
6
6
  Keywords: deep-learning,sparse-autoencoders,mechanistic-interpretability,PyTorch
{sae_lens-6.11.0.dist-info → sae_lens-6.12.0.dist-info}/RECORD RENAMED
@@ -1,4 +1,4 @@
1
- sae_lens/__init__.py,sha256=87qzm9CROR5ywD55DSxpt_ut24G4O8MR2zGYDtHEA0Y,3589
1
+ sae_lens/__init__.py,sha256=kTzPE14dVRpEIV0_yFC9whYCiJt6hoMQ33YhgwO83EI,3589
2
2
  sae_lens/analysis/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
3
3
  sae_lens/analysis/hooked_sae_transformer.py,sha256=vRu6JseH1lZaEeILD5bEkQEQ1wYHHDcxD-f2olKmE9Y,14275
4
4
  sae_lens/analysis/neuronpedia_integration.py,sha256=Gx1W7hUBEuMoasNcnOnZ1wmqbXDd1pSZ1nqKEya1HQc,4962
@@ -12,10 +12,10 @@ sae_lens/loading/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,
12
12
  sae_lens/loading/pretrained_sae_loaders.py,sha256=CVzHntSUKR1X3_gAqn8K_Ajq8D85qBrmrgEgU93IV4A,49609
13
13
  sae_lens/loading/pretrained_saes_directory.py,sha256=4Vn-Jex6SveD7EbxcSOBv8cx1gkPfUMLU1QOP-ww1ZE,3752
14
14
  sae_lens/pretokenize_runner.py,sha256=w0f6SfZLAxbp5eAAKnet8RqUB_DKofZ9RGsoJwFnYbA,7058
15
- sae_lens/pretrained_saes.yaml,sha256=d6FYfWTdVAPlOCM55C1ICS6lF9nWPPVNwjlXCa9p7NU,600468
15
+ sae_lens/pretrained_saes.yaml,sha256=n2BJVj8relA8OCbn9epJ5dFNg1_e9F9a4TouqjkXnts,602553
16
16
  sae_lens/registry.py,sha256=nhy7BPSudSATqW4lo9H_k3Na7sfGHmAf9v-3wpnLL_o,1490
17
17
  sae_lens/saes/__init__.py,sha256=jVwazK8Q6dW5J6_zFXPoNAuBvSxgziQ8eMOjGM3t-X8,1475
18
- sae_lens/saes/batchtopk_sae.py,sha256=CyaFG2hMyyDaEaXXrAMJC8wQDW1JoddTKF5mvxxBQKY,3395
18
+ sae_lens/saes/batchtopk_sae.py,sha256=GX_J0vH4vzeLqYxl0mkfsZQpFEoCEHMR4dIG8fz8N8w,3449
19
19
  sae_lens/saes/gated_sae.py,sha256=qcmM9JwBA8aZR8z_IRHV1_gQX-q_63tKewWXRnhdXuo,8986
20
20
  sae_lens/saes/jumprelu_sae.py,sha256=HHBF1sJ95lZvxwP5vwLSQFKdnJN2KKYK0WAEaLTrta0,13399
21
21
  sae_lens/saes/sae.py,sha256=gdUZuLaOHQrPjbDj-nZI813B6-_mNAnV9i9z4qTnpHk,38255
@@ -33,7 +33,7 @@ sae_lens/training/types.py,sha256=qSjmGzXf3MLalygG0psnVjmhX_mpLmL47MQtZfe7qxg,81
33
33
  sae_lens/training/upload_saes_to_huggingface.py,sha256=r_WzI1zLtGZ5TzAxuG3xa_8T09j3zXJrWd_vzPsPGkQ,4469
34
34
  sae_lens/tutorial/tsea.py,sha256=fd1am_XXsf2KMbByDapJo-2qlxduKaa62Z2qcQZ3QKU,18145
35
35
  sae_lens/util.py,sha256=lW7fBn_b8quvRYlen9PUmB7km60YhKyjmuelB1f6KzQ,2253
36
- sae_lens-6.11.0.dist-info/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
37
- sae_lens-6.11.0.dist-info/METADATA,sha256=yEZL6kzNBUKjzd0d73Hiup6SpLNtNEGUOgFu0MwgVpo,5245
38
- sae_lens-6.11.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
39
- sae_lens-6.11.0.dist-info/RECORD,,
36
+ sae_lens-6.12.0.dist-info/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
37
+ sae_lens-6.12.0.dist-info/METADATA,sha256=6L9rhKmaK8oQdRZ9-p0TxqpbBXCaY4bcwKqHP_NYnEw,5245
38
+ sae_lens-6.12.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
39
+ sae_lens-6.12.0.dist-info/RECORD,,