Add files using upload-large-folder tool
Browse files- README.md +53 -0
- attn_out_all/layer_0_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_0_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_10_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_10_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_11_width_262k_l0_big/config.json +10 -0
- attn_out_all/layer_12_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_12_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_12_width_262k_l0_big/config.json +10 -0
- attn_out_all/layer_12_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_13_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_13_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_14_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_16_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_16_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_17_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_18_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_1_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_20_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_20_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_21_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_22_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_22_width_262k_l0_big/config.json +10 -0
- attn_out_all/layer_23_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_23_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_24_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_24_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_26_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_26_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_27_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_27_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_28_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_28_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_29_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_29_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_30_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_30_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_31_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_31_width_262k_l0_big/config.json +10 -0
- attn_out_all/layer_32_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_32_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_33_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_3_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_4_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_5_width_16k_l0_big/config.json +10 -0
- attn_out_all/layer_5_width_262k_l0_big/config.json +10 -0
- attn_out_all/layer_6_width_262k_l0_small/config.json +10 -0
- attn_out_all/layer_7_width_262k_l0_big/config.json +10 -0
- attn_out_all/layer_8_width_16k_l0_small/config.json +10 -0
- attn_out_all/layer_8_width_262k_l0_big/config.json +10 -0
README.md
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
---
|
| 3 |
+
license: cc-by-4.0
|
| 4 |
+
library_name: saelens
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# 1. Gemma Scope 2
|
| 8 |
+
|
| 9 |
+
Gemma Scope 2 is a comprehensive, open suite of sparse autoencoders and transcoders for a range of model 4bs and versions in the Gemma 3 model family. We have SAEs on three different sites (as well as transcoders) for every layer of the pretrained and instruction-tuned models of parameter 4bs 270M, 1B, 4B, 12B and 27B. We also include several multi-layer SAE variants: partial residual stream crosscoders for every base Gemma 3 model, and cross-layer transcoders for the 270M and 1B models.
|
| 10 |
+
|
| 11 |
+
Sparse Autoencoders are a "microscope" of sorts that can help us break down a model's internal activations into the underlying concepts, just as biologists use microscopes to study the individual cells of plants and animals.
|
| 12 |
+
|
| 13 |
+
See our landing page for details on the whole suite.
|
| 14 |
+
|
| 15 |
+
# 2. What Is In This Repo?
|
| 16 |
+
|
| 17 |
+
This repo contains a specific set of SAEs and transcoders: the ones trained on Gemma V3 {4b_upper} {pt_upper}. Every folder here contains a different suite of models. Each of the folders in this page are named for the type of model that was trained:
|
| 18 |
+
|
| 19 |
+
- Single-layer models
|
| 20 |
+
- `resid_post`, `attn_out` and `mlp_out` contain SAEs at 4 different layers (25%, 50%, 65% and 85% depth) and a variety of widths & L0 values, trained on the model's residual stream, attention output, and MLP output respectively.
|
| 21 |
+
- `transcoder` contain a range of transcoders (or skip-transcoders) trained on the same 4 layers, with a variety of widths & L0 values.
|
| 22 |
+
- `resid_post_all`, `attn_out_all`, `mlp_out_all` and `transcoder_all` contain a smaller range of widths & L0 values, but for every single layer in the model.
|
| 23 |
+
- Multi-layer models
|
| 24 |
+
- `crosscoder` contains a set of weakly causal crosscoders which were trained on 4 concatenated layers of the residual stream (the same as those we trained our subsets on)
|
| 25 |
+
- `clt` contains a set of cross-layer transcoders, which were trained to reconstruct the whole model's MLP outputs from the residual stream values just before each MLP layer.
|
| 26 |
+
|
| 27 |
+
So for example, `google/gemma-scope-2-{4b}-{pt}/resid_post` contains a range of SAEs trained on the residual stream of `gemma-v3-270m-pt` at 4 different layers.
|
| 28 |
+
|
| 29 |
+
# 3. How can I use these SAEs straight away?
|
| 30 |
+
|
| 31 |
+
```py
|
| 32 |
+
from sae_lens import SAE # pip install sae-lens
|
| 33 |
+
|
| 34 |
+
sae, cfg_dict, sparsity = SAE.from_pretrained(
|
| 35 |
+
release = "gemma-scope-2-{4b}-{pt}-resid_post",
|
| 36 |
+
sae_id = "layer_12_width_16k_l0_small",
|
| 37 |
+
)
|
| 38 |
+
```
|
| 39 |
+
|
| 40 |
+
# 4. Which SAE should I use?
|
| 41 |
+
|
| 42 |
+
Unless you're doing full circuit-style analysis, we recommend using SAEs / transcoders from the layer subset folders, e.g. `resid_post` or `transcoder`. Assuming you're using residual stream SAEs from `resid_post`, then:
|
| 43 |
+
|
| 44 |
+
- **Width**: our SAEs have widths 16k, 64k, 256k, 1m. You can visit Neuronpedia to get a qualitative sense of what kinds of features you can find at different widths, but we generally recommend using 64k or 256k.
|
| 45 |
+
- **L0**: our SAEs have target L0 values "small" (10-20), "medium" (30-60) or "large" (60-150)". You can also look at the `config.json` file saved with every SAE's parameters to check exactly what the L0 is (or just visit the Neuronpedia page!). We generally recommend using "medium" which is useful for most tasks, although this might vary depending on your exact use case. Again you can visit Neuronpedia to get a sense of what kind of features each model type finds.
|
| 46 |
+
|
| 47 |
+
# 4. Point of Contact
|
| 48 |
+
|
| 49 |
+
Point of contact: Callum McDougall
|
| 50 |
+
Contact by email: [email protected]
|
| 51 |
+
|
| 52 |
+
# 5. Citation
|
| 53 |
+
Paper: (link to go here)
|
attn_out_all/layer_0_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.0.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.0.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 60,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_0_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.0.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.0.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 10,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_10_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.10.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.10.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 112,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_10_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.10.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.10.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 18,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_11_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.11.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.11.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 118,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_12_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.12.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.12.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_12_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.12.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.12.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_12_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.12.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.12.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_12_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.12.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.12.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_13_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.13.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.13.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_13_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.13.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.13.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_14_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.14.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.14.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_16_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.16.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.16.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_16_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.16.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.16.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_17_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.17.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.17.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_18_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.18.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.18.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_1_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.1.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.1.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 10,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_20_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.20.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.20.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_20_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.20.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.20.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_21_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.21.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.21.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_22_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.22.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.22.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_22_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.22.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.22.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_23_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.23.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.23.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_23_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.23.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.23.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_24_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.24.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.24.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_24_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.24.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.24.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_26_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.26.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.26.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_26_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.26.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.26.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_27_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.27.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.27.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_27_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.27.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.27.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_28_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.28.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.28.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_28_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.28.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.28.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_29_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.29.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.29.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_29_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.29.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.29.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_30_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.30.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.30.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_30_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.30.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.30.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_31_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.31.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.31.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_31_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.31.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.31.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 120,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_32_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.32.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.32.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_32_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.32.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.32.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_33_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.33.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.33.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 20,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_3_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.3.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.3.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 12,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_4_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.4.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.4.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 81,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_5_width_16k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.5.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.5.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 86,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_5_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.5.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.5.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 86,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_6_width_262k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.6.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.6.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 15,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_7_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.7.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.7.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 97,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_8_width_16k_l0_small/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.8.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.8.self_attn.o_proj.input",
|
| 4 |
+
"width": 16384,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 17,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|
attn_out_all/layer_8_width_262k_l0_big/config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hf_hook_point_in": "model.layers.8.self_attn.o_proj.input",
|
| 3 |
+
"hf_hook_point_out": "model.layers.8.self_attn.o_proj.input",
|
| 4 |
+
"width": 262144,
|
| 5 |
+
"model_name": "google/gemma-3-4b-pt",
|
| 6 |
+
"architecture": "jump_relu",
|
| 7 |
+
"l0": 102,
|
| 8 |
+
"affine_connection": false,
|
| 9 |
+
"type": "sae"
|
| 10 |
+
}
|