Instructions to use google/gemma-scope-2-27b-pt with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- SAELens
How to use google/gemma-scope-2-27b-pt with SAELens:
# pip install sae-lens from sae_lens import SAE sae, cfg_dict, sparsity = SAE.from_pretrained( release = "RELEASE_ID", # e.g., "gpt2-small-res-jb". See other options in https://github.com/jbloomAus/SAELens/blob/main/sae_lens/pretrained_saes.yaml sae_id = "SAE_ID", # e.g., "blocks.8.hook_resid_pre". Won't always be a hook point ) - Notebooks
- Google Colab
- Kaggle
Add files using upload-large-folder tool
Browse files- transcoder_all/layer_14_width_262k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_17_width_16k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_1_width_16k_l0_small/config.json +9 -0
- transcoder_all/layer_1_width_16k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_20_width_16k_l0_big/config.json +9 -0
- transcoder_all/layer_20_width_16k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_21_width_262k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_23_width_262k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_24_width_16k_l0_small/config.json +9 -0
- transcoder_all/layer_25_width_262k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_26_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_28_width_16k_l0_big/config.json +9 -0
- transcoder_all/layer_30_width_16k_l0_small/config.json +9 -0
- transcoder_all/layer_31_width_16k_l0_small/config.json +9 -0
- transcoder_all/layer_32_width_262k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_33_width_262k_l0_big/config.json +9 -0
- transcoder_all/layer_35_width_262k_l0_big/config.json +9 -0
- transcoder_all/layer_36_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_39_width_262k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_3_width_16k_l0_big/config.json +9 -0
- transcoder_all/layer_3_width_16k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_41_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_41_width_262k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_45_width_262k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_45_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_46_width_262k_l0_big/config.json +9 -0
- transcoder_all/layer_47_width_16k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_47_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_49_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_4_width_16k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_51_width_16k_l0_small/config.json +9 -0
- transcoder_all/layer_51_width_16k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_51_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_52_width_16k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_54_width_262k_l0_small/config.json +9 -0
- transcoder_all/layer_56_width_16k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_58_width_16k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_58_width_262k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_59_width_262k_l0_big_affine/config.json +9 -0
- transcoder_all/layer_5_width_16k_l0_big/config.json +9 -0
- transcoder_all/layer_60_width_262k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_6_width_262k_l0_small_affine/config.json +9 -0
- transcoder_all/layer_7_width_16k_l0_small/config.json +9 -0
transcoder_all/layer_14_width_262k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.14.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.14.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 16,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_17_width_16k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.17.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.17.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 18,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_1_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.1.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.1.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 10,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_1_width_16k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.1.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.1.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 10,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_20_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.20.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.20.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 118,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_20_width_16k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.20.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.20.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 118,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_21_width_262k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.21.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.21.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_23_width_262k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.23.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.23.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_24_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.24.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.24.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_25_width_262k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.25.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.25.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_26_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.26.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.26.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_28_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.28.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.28.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_30_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.30.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.30.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_31_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.31.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.31.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_32_width_262k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.32.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.32.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_33_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.33.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.33.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_35_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.35.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.35.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_36_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.36.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.36.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_39_width_262k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.39.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.39.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_3_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.3.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.3.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 68,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_3_width_16k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.3.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.3.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 68,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_41_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.41.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.41.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_41_width_262k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.41.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.41.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_45_width_262k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.45.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.45.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_45_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.45.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.45.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_46_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.46.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.46.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_47_width_16k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.47.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.47.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_47_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.47.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.47.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_49_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.49.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.49.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_4_width_16k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.4.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.4.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 71,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_51_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.51.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.51.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_51_width_16k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.51.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.51.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_51_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.51.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.51.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_52_width_16k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.52.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.52.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_54_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.54.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.54.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_56_width_16k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.56.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.56.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_58_width_16k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.58.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.58.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_58_width_262k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.58.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.58.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_59_width_262k_l0_big_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.59.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.59.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_5_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.5.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.5.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 74,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|
transcoder_all/layer_60_width_262k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.60.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.60.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_6_width_262k_l0_small_affine/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.6.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.6.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 12,
|
| 8 |
+
"affine_connection": true
|
| 9 |
+
}
|
transcoder_all/layer_7_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.7.pre_feedforward_layernorm.output",
|
| 3 |
+
"hf_hook_point_out": "model.layers.7.post_feedforward_layernorm.output",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "gemma-v3-27b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 13,
|
| 8 |
+
"affine_connection": false
|
| 9 |
+
}
|