Hugging Face's logo

Datasets:
juiceb0xc0de
/
gemma-4-e2b-it-SAE

Tasks:
Feature Extraction
Size:
10B<n<100B
Tags:
sparse-autoencoder
mechanistic-interpretability
sae
gemma
brain-atlas
jumprelu
License:
Dataset card Files
xet
 Community
1
juiceb0xc0de commited on
Commit
e1a1b62
·
verified ·
1 Parent(s): 4e0fba2

Copy SAE layer 28 from model repo

Browse files
Files changed (3) hide show
  1. layer_28_s0/checkpoint_full.pt +3 -0
  2. layer_28_s0/meta.json +117 -0
  3. layer_28_s0/sae.pt +3 -0
layer_28_s0/checkpoint_full.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ecf1edfd94e842bab281f676cfcc6d3c2a3864ca1971206316007411807f1f5c
3
+ size 1813939659
layer_28_s0/meta.json ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "layer": 28,
3
+ "seed": 0,
4
+ "model_id": "google/gemma-4-E2B-it",
5
+ "d_in": 1536,
6
+ "n_features": 49152,
7
+ "k": 50,
8
+ "batch_tokens": 32768,
9
+ "n_steps": 2001,
10
+ "lr": 0.0002,
11
+ "lambda_l0": 1.0146211969683001e-06,
12
+ "tier": "mid",
13
+ "preflight": {
14
+ "activation_norm_probe": 2.982285261154175,
15
+ "activation_norm_ref": 0.7526426911354065,
16
+ "initial_l0_probe": 71.95849609375,
17
+ "initial_lr_multiplier": 1.65,
18
+ "early_pulse_multiplier": 1.11513359375,
19
+ "layer_lr_multiplier": 1.27
20
+ },
21
+ "scheduler_mode": "BASELINE",
22
+ "scheduler_phase": "DESCENT",
23
+ "scheduler_phase_summary": {
24
+ "phase": "DESCENT",
25
+ "phase_step": 2000,
26
+ "pin_entry_step": null,
27
+ "pinned_lambda": null,
28
+ "pin_ev_count": 0,
29
+ "pin_retry_count": 0
30
+ },
31
+ "scheduler_transitions": 0,
32
+ "total_tokens": 65568768,
33
+ "early_stopped": true,
34
+ "path": "rolling",
35
+ "best_ev": 0.9425579824331939,
36
+ "best_ev_step": 2000,
37
+ "final_metrics": {
38
+ "recon_loss": 0.474029,
39
+ "mean_l0": 49.5,
40
+ "dead_pct": 0.0,
41
+ "resampled": 0,
42
+ "ev": 0.9438821718753864,
43
+ "nonlinear_err": null,
44
+ "linear_err": null,
45
+ "qo_gap": 0.729,
46
+ "activation_norm_probe": 2.982285261154175,
47
+ "activation_norm_ref": 0.7526426911354065,
48
+ "initial_l0_probe": 71.95849609375,
49
+ "initial_lr_multiplier": 1.65,
50
+ "early_pulse_multiplier": 1.11513359375,
51
+ "layer_lr_multiplier": 1.27
52
+ },
53
+ "training_curve": {
54
+ "recon_loss": [
55
+ 0.915491,
56
+ 0.662518,
57
+ 0.60314,
58
+ 0.572743,
59
+ 0.523826,
60
+ 0.468072,
61
+ 0.460126,
62
+ 0.474029
63
+ ],
64
+ "mean_l0": [
65
+ 23.52,
66
+ 25.14,
67
+ 26.01,
68
+ 29.26,
69
+ 32.46,
70
+ 38.76,
71
+ 44.4,
72
+ 49.5
73
+ ],
74
+ "dead_pct": [
75
+ 0.0,
76
+ 0.0,
77
+ 0.0,
78
+ 0.0,
79
+ 0.0,
80
+ 0.0,
81
+ 0.0,
82
+ 0.0
83
+ ],
84
+ "resampled": [
85
+ 0,
86
+ 0,
87
+ 0,
88
+ 0,
89
+ 0,
90
+ 0,
91
+ 0,
92
+ 0
93
+ ],
94
+ "ev": [
95
+ 0.8912912938290513,
96
+ 0.9191840790466639,
97
+ 0.9252288325698044,
98
+ 0.93115186118049,
99
+ 0.9355976847491873,
100
+ 0.9456808285407405,
101
+ 0.9442700293494312,
102
+ 0.9438821718753864
103
+ ],
104
+ "nonlinear_err": [],
105
+ "linear_err": [],
106
+ "qo_gap": [
107
+ 0.8181,
108
+ 0.7608,
109
+ 0.7542,
110
+ 0.7478,
111
+ 0.7418,
112
+ 0.7422,
113
+ 0.7308,
114
+ 0.729
115
+ ]
116
+ }
117
+ }
layer_28_s0/sae.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b8d78c3771fc288a97dffbfa74f91efbf537ee30d5b82b73d05595744d8d596a
3
+ size 604381549