jjho-choi commited on
Commit
005a77d
·
verified ·
1 Parent(s): 7f6b0b4

Add files using upload-large-folder tool

Browse files
Files changed (50) hide show
  1. README.md +60 -0
  2. vision/dinov2-b_layer10/consdl/x16/l0_128/code_metrics_step_final.json +136 -0
  3. vision/dinov2-b_layer10/consdl/x16/l0_128/dict_metrics_step_final.json +550 -0
  4. vision/dinov2-b_layer10/consdl/x16/l0_128/seed0/config.json +24 -0
  5. vision/dinov2-b_layer10/consdl/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  6. vision/dinov2-b_layer10/consdl/x16/l0_128/seed1/config.json +24 -0
  7. vision/dinov2-b_layer10/consdl/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  8. vision/dinov2-b_layer10/consdl/x16/l0_128/seed2/config.json +24 -0
  9. vision/dinov2-b_layer10/consdl/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  10. vision/dinov2-b_layer10/consdl/x16/l0_32/code_metrics_step_final.json +136 -0
  11. vision/dinov2-b_layer10/consdl/x16/l0_32/dict_metrics_step_final.json +550 -0
  12. vision/dinov2-b_layer10/consdl/x16/l0_32/seed0/config.json +24 -0
  13. vision/dinov2-b_layer10/consdl/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  14. vision/dinov2-b_layer10/consdl/x16/l0_32/seed1/config.json +24 -0
  15. vision/dinov2-b_layer10/consdl/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  16. vision/dinov2-b_layer10/consdl/x16/l0_32/seed2/config.json +24 -0
  17. vision/dinov2-b_layer10/consdl/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  18. vision/dinov2-b_layer10/consdl/x16/l0_64/code_metrics_step_final.json +136 -0
  19. vision/dinov2-b_layer10/consdl/x16/l0_64/dict_metrics_step_final.json +550 -0
  20. vision/dinov2-b_layer10/consdl/x16/l0_64/seed0/config.json +24 -0
  21. vision/dinov2-b_layer10/consdl/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  22. vision/dinov2-b_layer10/consdl/x16/l0_64/seed1/config.json +24 -0
  23. vision/dinov2-b_layer10/consdl/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  24. vision/dinov2-b_layer10/consdl/x16/l0_64/seed2/config.json +24 -0
  25. vision/dinov2-b_layer10/consdl/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  26. vision/dinov2-b_layer10/consdl/x4/l0_128/code_metrics_step_final.json +136 -0
  27. vision/dinov2-b_layer10/consdl/x4/l0_128/dict_metrics_step_final.json +595 -0
  28. vision/dinov2-b_layer10/consdl/x4/l0_128/seed0/config.json +24 -0
  29. vision/dinov2-b_layer10/consdl/x4/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  30. vision/dinov2-b_layer10/consdl/x4/l0_128/seed1/config.json +24 -0
  31. vision/dinov2-b_layer10/consdl/x4/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  32. vision/dinov2-b_layer10/consdl/x4/l0_128/seed2/config.json +24 -0
  33. vision/dinov2-b_layer10/consdl/x4/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  34. vision/dinov2-b_layer10/consdl/x4/l0_32/code_metrics_step_final.json +136 -0
  35. vision/dinov2-b_layer10/consdl/x4/l0_32/dict_metrics_step_final.json +550 -0
  36. vision/dinov2-b_layer10/consdl/x4/l0_32/seed0/config.json +24 -0
  37. vision/dinov2-b_layer10/consdl/x4/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  38. vision/dinov2-b_layer10/consdl/x4/l0_32/seed1/config.json +24 -0
  39. vision/dinov2-b_layer10/consdl/x4/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  40. vision/dinov2-b_layer10/consdl/x4/l0_32/seed2/config.json +24 -0
  41. vision/dinov2-b_layer10/consdl/x4/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  42. vision/dinov2-b_layer10/consdl/x4/l0_64/code_metrics_step_final.json +136 -0
  43. vision/dinov2-b_layer10/consdl/x4/l0_64/dict_metrics_step_final.json +595 -0
  44. vision/dinov2-b_layer10/consdl/x4/l0_64/seed0/config.json +24 -0
  45. vision/dinov2-b_layer10/consdl/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  46. vision/dinov2-b_layer10/consdl/x4/l0_64/seed1/config.json +24 -0
  47. vision/dinov2-b_layer10/consdl/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  48. vision/dinov2-b_layer10/consdl/x4/l0_64/seed2/config.json +24 -0
  49. vision/dinov2-b_layer10/consdl/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  50. vision/dinov2-b_layer10/consdl/x8/l0_64/code_metrics_step_final.json +136 -0
README.md ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ tags:
4
+ - sparse-autoencoder
5
+ - dictionary-learning
6
+ - interpretability
7
+ ---
8
+
9
+ # ConSDL checkpoints
10
+
11
+ Model checkpoints for the vision and language experiments in *ConSDL:
12
+ Consistent Sparse Dictionary Learning*. This collection includes ConSDL
13
+ and six SAE baselines across four backbones, three dictionary sizes, and
14
+ three sparsity levels, with three training seeds per configuration.
15
+
16
+ ## Layout
17
+
18
+ ```
19
+ <modality>/<backbone>_layer<L>/<arch>/x<E>/<level>/seed<S>/
20
+ config.json
21
+ model.safetensors
22
+ stats.<dataset>.split_<split>.seed_no/ per-latent firing statistics
23
+ <modality>/<backbone>_layer<L>/<arch>/x<E>/<level>/
24
+ dict_metrics_step_final.json cross-seed dictionary metrics (DCS, #dead, #dup)
25
+ code_metrics_step_final.json cross-seed code metrics (MCC, R²)
26
+ ```
27
+
28
+ | field | values |
29
+ |---|---|
30
+ | modality / backbone | `vision/clip-vit-b-16_layer10`, `vision/dinov2-b_layer10`, `lm/pythia-160m_layer8`, `lm/gpt2-small_layer8` |
31
+ | arch | `consdl`, `relu`, `topk`, `batchtopk`, `matryoshka`, `mp`, `isae` |
32
+ | x\<E\> | dictionary size = E × 768: `x4`, `x8`, `x16` |
33
+ | level | target L0: `l0_32`, `l0_64`, `l0_128` |
34
+ | seed | `seed0`, `seed1`, `seed2` |
35
+
36
+ Vision SAEs are trained on ImageNet patch tokens (one epoch), language SAEs on
37
+ 500M FineWeb tokens (residual stream after the given block). The exact
38
+ hyper-parameters of every checkpoint are in its `config.json`; the
39
+ per-cell values are also listed in the paper's appendix.
40
+
41
+ ## Usage
42
+
43
+ ```python
44
+ from consdl.io.hub import load_pretrained
45
+
46
+ sae = load_pretrained("vision/clip-vit-b-16_layer10/consdl/x16/l0_64/seed0")
47
+ codes = sae.encode(activations) # (..., 768) -> (..., 12288)
48
+ recon = sae.decode(codes)
49
+ ```
50
+
51
+ To recompute the cross-seed metrics of a cell, download it and run the
52
+ metric CLIs on it:
53
+
54
+ ```python
55
+ from consdl.io.hub import download
56
+ cell = download("vision/clip-vit-b-16_layer10/consdl/x16/l0_64")
57
+ ```
58
+ ```bash
59
+ python -m consdl.cli.dict_metrics --tag <cell>
60
+ ```
vision/dinov2-b_layer10/consdl/x16/l0_128/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 136.76897016861219,
20
+ "recon_r2": 0.9546725109176148,
21
+ "l1_per_token": 57.312201673617835
22
+ },
23
+ "seed1": {
24
+ "l0": 137.58134370946823,
25
+ "recon_r2": 0.9546184225558791,
26
+ "l1_per_token": 57.410340810621804
27
+ },
28
+ "seed2": {
29
+ "l0": 137.07644747081713,
30
+ "recon_r2": 0.9546766448861698,
31
+ "l1_per_token": 57.39469719095013
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.7694935938034413,
38
+ "mcc_abs": 0.7694936038875752,
39
+ "act_iou": 0.43669568523156105,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9985003921725566,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.793687663114197,
44
+ "0.7": 0.8283128829822111,
45
+ "0.8": 0.8818315588492487,
46
+ "0.9": 0.9439464145227352
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.7935032503978801,
50
+ "0.7": 0.8281970408944044,
51
+ "0.8": 0.8818518484377734,
52
+ "0.9": 0.943949992495095
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.769911134415597,
58
+ "mcc_abs": 0.7699119086685071,
59
+ "act_iou": 0.43513276290015757,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.9984944520726232,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.7914320355669494,
64
+ "0.7": 0.8259946188656512,
65
+ "0.8": 0.880308997462409,
66
+ "0.9": 0.943378116167842
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.7916717795931214,
70
+ "0.7": 0.8259684697474656,
71
+ "0.8": 0.8803866054816267,
72
+ "0.9": 0.9433840904939271
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.7714201854727096,
78
+ "mcc_abs": 0.7714201854727096,
79
+ "act_iou": 0.4383595450919215,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.9985040568541657,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.795177450907253,
84
+ "0.7": 0.8287873831626962,
85
+ "0.8": 0.8826430105415409,
86
+ "0.9": 0.9427993392858862
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.7944126443425757,
90
+ "0.7": 0.8285128617197115,
91
+ "0.8": 0.8828368909763905,
92
+ "0.9": 0.9427947111105671
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 137.14225378296587,
106
+ "l0_std": 0.4101652660936898,
107
+ "recon_r2_mean": 0.9546558594532213,
108
+ "recon_r2_std": 3.248712646504796e-05,
109
+ "l1_per_token_mean": 57.37241322506326,
110
+ "l1_per_token_std": 0.052728113039502124,
111
+ "input_r2_mean": 0.9984996336997819,
112
+ "input_r2_std": 4.8471040665580235e-06,
113
+ "mcc_signed_mean": 0.7702749712305826,
114
+ "mcc_signed_std": 0.0010135195044880558,
115
+ "mcc_abs_mean": 0.770275232676264,
116
+ "mcc_abs_std": 0.0010133767325156992,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.7934323831961332,
119
+ "0.7": 0.8276982950035195,
120
+ "0.8": 0.8815945222843995,
121
+ "0.9": 0.9433746233254879
122
+ },
123
+ "matched_mcc_signed_band": 0.8615249559523851,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.7931958914445257,
126
+ "0.7": 0.8275594574538605,
127
+ "0.8": 0.8816917816319303,
128
+ "0.9": 0.9433762646998631
129
+ },
130
+ "matched_mcc_abs_band": 0.8614558488075449,
131
+ "act_iou_mean": 0.43672933107454676,
132
+ "act_iou_std": 0.0016136541947808572,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/consdl/x16/l0_128/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": 138,
22
+ "l0_raw": 137.6256,
23
+ "recon_r2": 0.9547,
24
+ "l1_per_token": 57.3651,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 10346,
34
+ "0.6": 91,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 3898,
41
+ "0.6": 39,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 8,
48
+ "0.6": 3,
49
+ "0.7": 1,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.5215243697166443,
54
+ "n_dup_abs": {
55
+ "0.5": 10344,
56
+ "0.6": 113,
57
+ "0.7": 4,
58
+ "0.8": 0,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 3900,
63
+ "0.6": 50,
64
+ "0.7": 2,
65
+ "0.8": 0,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 9,
70
+ "0.6": 3,
71
+ "0.7": 2,
72
+ "0.8": 1,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.5223107933998108
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 138,
81
+ "l0_raw": 137.6256,
82
+ "recon_r2": 0.9546,
83
+ "l1_per_token": 57.4629,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 0,
89
+ "dense_rate": 0.0,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 10399,
93
+ "0.6": 109,
94
+ "0.7": 2,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 3880,
100
+ "0.6": 44,
101
+ "0.7": 1,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 9,
107
+ "0.6": 4,
108
+ "0.7": 2,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.5218309760093689,
113
+ "n_dup_abs": {
114
+ "0.5": 10415,
115
+ "0.6": 135,
116
+ "0.7": 8,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 3893,
122
+ "0.6": 57,
123
+ "0.7": 4,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 9,
129
+ "0.6": 4,
130
+ "0.7": 2,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.5226430296897888
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 138,
140
+ "l0_raw": 137.6256,
141
+ "recon_r2": 0.9547,
142
+ "l1_per_token": 57.4464,
143
+ "stats_source": "stats",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 0,
148
+ "dense_rate": 0.0,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 10420,
152
+ "0.6": 83,
153
+ "0.7": 0,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 3890,
159
+ "0.6": 34,
160
+ "0.7": 0,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 8,
166
+ "0.6": 4,
167
+ "0.7": 1,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.521738588809967,
172
+ "n_dup_abs": {
173
+ "0.5": 10419,
174
+ "0.6": 111,
175
+ "0.7": 6,
176
+ "0.8": 0,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 3890,
181
+ "0.6": 48,
182
+ "0.7": 3,
183
+ "0.8": 0,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 8,
188
+ "0.6": 4,
189
+ "0.7": 2,
190
+ "0.8": 1,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.5227003693580627
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.8129798769950867,
199
+ "dcs_abs": 0.8129990696907043,
200
+ "frac_sign_flipped": 0.0006510416666666666,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 12288,
204
+ "n_alive_b": 12288,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.8402630686759949,
207
+ "0.6": 0.8129270672798157,
208
+ "0.7": 0.8129801154136658,
209
+ "0.8": 0.8129798769950867,
210
+ "0.9": 0.8129798769950867
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.4723714192708333,
214
+ "0.6": 0.9952392578125,
215
+ "0.7": 0.9999593098958334,
216
+ "0.8": 1.0,
217
+ "0.9": 1.0
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.39691625831134536,
221
+ "0.6": 0.809056931095256,
222
+ "0.7": 0.8129470351680842,
223
+ "0.8": 0.8129798769950867,
224
+ "0.9": 0.8129798769950867
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.4580078125,
228
+ "0.6": 0.929443359375,
229
+ "0.7": 0.8131510416666666,
230
+ "0.8": 0.5865885416666666,
231
+ "0.9": 0.29638671875
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.8407085537910461,
235
+ "0.6": 0.813060998916626,
236
+ "0.7": 0.8130369186401367,
237
+ "0.8": 0.8129990696907043,
238
+ "0.9": 0.8129990696907043
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.472412109375,
242
+ "0.6": 0.9942626953125,
243
+ "0.7": 0.999755859375,
244
+ "0.8": 1.0,
245
+ "0.9": 1.0
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.39716090126603376,
249
+ "0.6": 0.8083962202363182,
250
+ "0.7": 0.8128384232986718,
251
+ "0.8": 0.8129990696907043,
252
+ "0.9": 0.8129990696907043
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.4584147135416667,
256
+ "0.6": 0.9285481770833334,
257
+ "0.7": 0.8131510416666666,
258
+ "0.8": 0.58642578125,
259
+ "0.9": 0.2964680989583333
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.8128573894500732,
264
+ "dcs_abs": 0.8128764629364014,
265
+ "frac_sign_flipped": 0.0008951822916666666,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 12288,
269
+ "n_alive_b": 12288,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.8401245474815369,
272
+ "0.6": 0.8120025992393494,
273
+ "0.7": 0.8128573894500732,
274
+ "0.8": 0.8128573894500732,
275
+ "0.9": 0.8128573894500732
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.471923828125,
279
+ "0.6": 0.9958902994791666,
280
+ "0.7": 1.0,
281
+ "0.8": 1.0,
282
+ "0.9": 1.0
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.3964747925492702,
286
+ "0.6": 0.8086655117343373,
287
+ "0.7": 0.8128573894500732,
288
+ "0.8": 0.8128573894500732,
289
+ "0.9": 0.8128573894500732
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.45849609375,
293
+ "0.6": 0.93505859375,
294
+ "0.7": 0.8168131510416666,
295
+ "0.8": 0.5861002604166666,
296
+ "0.9": 0.2902018229166667
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.8401395082473755,
300
+ "0.6": 0.8119817972183228,
301
+ "0.7": 0.8129231333732605,
302
+ "0.8": 0.8128764629364014,
303
+ "0.9": 0.8128764629364014
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.4721272786458333,
307
+ "0.6": 0.994873046875,
308
+ "0.7": 0.9997965494791666,
309
+ "0.8": 1.0,
310
+ "0.9": 1.0
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.396652779711682,
314
+ "0.6": 0.8078188046056312,
315
+ "0.7": 0.8127577437383782,
316
+ "0.8": 0.8128764629364014,
317
+ "0.9": 0.8128764629364014
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.4584147135416667,
321
+ "0.6": 0.9340006510416666,
322
+ "0.7": 0.81640625,
323
+ "0.8": 0.5861002604166666,
324
+ "0.9": 0.2902018229166667
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.8139712810516357,
329
+ "dcs_abs": 0.8139920234680176,
330
+ "frac_sign_flipped": 0.00048828125,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 12288,
334
+ "n_alive_b": 12288,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.8382914066314697,
337
+ "0.6": 0.8136950731277466,
338
+ "0.7": 0.8140022158622742,
339
+ "0.8": 0.8139712810516357,
340
+ "0.9": 0.8139712810516357
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.4690348307291667,
344
+ "0.6": 0.995361328125,
345
+ "0.7": 0.9999593098958334,
346
+ "0.8": 1.0,
347
+ "0.9": 1.0
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.39318786801110645,
351
+ "0.6": 0.8099206086772028,
352
+ "0.7": 0.8139690940273189,
353
+ "0.8": 0.8139712810516357,
354
+ "0.9": 0.8139712810516357
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.4562174479166667,
358
+ "0.6": 0.9293619791666666,
359
+ "0.7": 0.8150227864583334,
360
+ "0.8": 0.589111328125,
361
+ "0.9": 0.3025716145833333
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.8382344245910645,
365
+ "0.6": 0.8137317299842834,
366
+ "0.7": 0.8140093684196472,
367
+ "0.8": 0.8139920234680176,
368
+ "0.9": 0.8139920234680176
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.4689534505208333,
372
+ "0.6": 0.9942626953125,
373
+ "0.7": 0.9997151692708334,
374
+ "0.8": 1.0,
375
+ "0.9": 1.0
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.3930929257573249,
379
+ "0.6": 0.8090631031154771,
380
+ "0.7": 0.8137775135376918,
381
+ "0.8": 0.8139920234680176,
382
+ "0.9": 0.8139920234680176
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.4559733072916667,
386
+ "0.6": 0.9297688802083334,
387
+ "0.7": 0.8152669270833334,
388
+ "0.8": 0.5879720052083334,
389
+ "0.9": 0.3023274739583333
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 0,
403
+ "n_dead_std": 0.0,
404
+ "n_dense_mean": 0,
405
+ "n_dense_std": 0.0,
406
+ "l0_mean": 138,
407
+ "l0_std": 0.0,
408
+ "recon_r2_mean": 0.9546666666666667,
409
+ "recon_r2_std": 5.773502691895622e-05,
410
+ "l1_per_token_mean": 57.424800000000005,
411
+ "l1_per_token_std": 0.05235580197074583,
412
+ "babel_mu1_signed_mean": 0.52169797817866,
413
+ "babel_mu1_signed_std": 0.00015728563893739622,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 10388.333333333334,
416
+ "0.6": 94.33333333333333,
417
+ "0.7": 0.6666666666666666,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 23.75,
422
+ "n_dup_signed_band_std": 3.605551275463989,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 3889.3333333333335,
425
+ "0.6": 39.0,
426
+ "0.7": 0.3333333333333333,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 9.833333333333334,
431
+ "n_dup_clusters_signed_band_std": 1.3768926368215255,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 8.333333333333334,
434
+ "0.6": 3.6666666666666665,
435
+ "0.7": 1.3333333333333333,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 1.75,
440
+ "largest_dup_signed_band_std": 0.25,
441
+ "dcs_signed_mean": 0.8132695158322653,
442
+ "dcs_signed_std": 0.0006108245386438515,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.395526306290574,
445
+ "0.6": 0.8092143505022653,
446
+ "0.7": 0.8132578395484922,
447
+ "0.8": 0.8132695158322653,
448
+ "0.9": 0.8132695158322653
449
+ },
450
+ "dedup_dcs_signed_band": 0.8122528054288221,
451
+ "dedup_dcs_signed_band_std": 0.0006174795886841092,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.8395596742630005,
454
+ "0.6": 0.8128749132156372,
455
+ "0.7": 0.813279906908671,
456
+ "0.8": 0.8132695158322653,
457
+ "0.9": 0.8132695158322653
458
+ },
459
+ "matched_cos_signed_band": 0.8131734629472097,
460
+ "matched_cos_signed_band_std": 0.000657961300548334,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.4711100260416667,
463
+ "0.6": 0.9954969618055555,
464
+ "0.7": 0.999972873263889,
465
+ "0.8": 1.0,
466
+ "0.9": 1.0
467
+ },
468
+ "unique_frac_signed_band": 0.9988674587673612,
469
+ "unique_frac_signed_band_std": 9.230317917758508e-05,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.45757378472222227,
472
+ "0.6": 0.9312879774305555,
473
+ "0.7": 0.8149956597222222,
474
+ "0.8": 0.5872667100694444,
475
+ "0.9": 0.29638671875
476
+ },
477
+ "matched_frac_signed_band": 0.6574842664930556,
478
+ "matched_frac_signed_band_std": 0.0013666565686657288,
479
+ "babel_mu1_abs_mean": 0.5225513974825541,
480
+ "babel_mu1_abs_std": 0.00021033236285462398,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 10392.666666666666,
483
+ "0.6": 119.66666666666667,
484
+ "0.7": 6.0,
485
+ "0.8": 0.0,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 31.416666666666668,
489
+ "n_dup_abs_band_std": 3.7527767497325675,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 3894.3333333333335,
492
+ "0.6": 51.666666666666664,
493
+ "0.7": 3.0,
494
+ "0.8": 0.0,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 13.666666666666666,
498
+ "n_dup_clusters_abs_band_std": 1.3768926368215255,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 8.666666666666666,
501
+ "0.6": 3.6666666666666665,
502
+ "0.7": 2.0,
503
+ "0.8": 1.0,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 1.9166666666666665,
507
+ "largest_dup_abs_band_std": 0.14433756729740646,
508
+ "dcs_abs_mean": 0.8132891853650411,
509
+ "dcs_abs_std": 0.0006117549781545401,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.3956355355783469,
512
+ "0.6": 0.8084260426524755,
513
+ "0.7": 0.8131245601915805,
514
+ "0.8": 0.8132891853650411,
515
+ "0.9": 0.8132891853650411
516
+ },
517
+ "dedup_dcs_abs_band": 0.8120322433935346,
518
+ "dedup_dcs_abs_band_std": 0.0005944561666284777,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.8396941622098287,
521
+ "0.6": 0.812924842039744,
522
+ "0.7": 0.8133231401443481,
523
+ "0.8": 0.8132891853650411,
524
+ "0.9": 0.8132891853650411
525
+ },
526
+ "matched_cos_abs_band": 0.8132065882285436,
527
+ "matched_cos_abs_band_std": 0.0006528472876833653,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.47116427951388884,
530
+ "0.6": 0.9944661458333334,
531
+ "0.7": 0.999755859375,
532
+ "0.8": 1.0,
533
+ "0.9": 1.0
534
+ },
535
+ "unique_frac_abs_band": 0.9985555013020834,
536
+ "unique_frac_abs_band_std": 9.703971368577286e-05,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.4576009114583333,
539
+ "0.6": 0.9307725694444445,
540
+ "0.7": 0.81494140625,
541
+ "0.8": 0.5868326822916666,
542
+ "0.9": 0.29633246527777773
543
+ },
544
+ "matched_frac_abs_band": 0.6572197808159722,
545
+ "matched_frac_abs_band_std": 0.0014226027021161617,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 12288
549
+ }
550
+ }
vision/dinov2-b_layer10/consdl/x16/l0_128/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.6,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 5,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1675,
3
+ "r2_score": 0.9547,
4
+ "l0_per_token_0": 0.0112,
5
+ "l0_per_token_0.001": 0.0112,
6
+ "l0_per_token_0.01": 0.0111,
7
+ "l0_per_token_0.05": 0.0099,
8
+ "l0_per_token_0.1": 0.0082,
9
+ "l0_per_token_0.5": 0.0026,
10
+ "l0_per_token_1.0": 0.001,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 57.3651
15
+ }
vision/dinov2-b_layer10/consdl/x16/l0_128/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.6,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 5,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1676,
3
+ "r2_score": 0.9546,
4
+ "l0_per_token_0": 0.0112,
5
+ "l0_per_token_0.001": 0.0112,
6
+ "l0_per_token_0.01": 0.0112,
7
+ "l0_per_token_0.05": 0.01,
8
+ "l0_per_token_0.1": 0.0082,
9
+ "l0_per_token_0.5": 0.0026,
10
+ "l0_per_token_1.0": 0.001,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 57.4629
15
+ }
vision/dinov2-b_layer10/consdl/x16/l0_128/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.6,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 5,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1674,
3
+ "r2_score": 0.9547,
4
+ "l0_per_token_0": 0.0112,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 57.4464
9
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 35.1568093385214,
20
+ "recon_r2": 0.9498433542063758,
21
+ "l1_per_token": 50.521696527612356
22
+ },
23
+ "seed1": {
24
+ "l0": 35.39899610894942,
25
+ "recon_r2": 0.9498487807206017,
26
+ "l1_per_token": 50.470284029820775
27
+ },
28
+ "seed2": {
29
+ "l0": 35.15252399481193,
30
+ "recon_r2": 0.9498539621813498,
31
+ "l1_per_token": 50.46092977149795
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.6637414743484052,
38
+ "mcc_abs": 0.6637435681283992,
39
+ "act_iou": 0.43222036387868945,
40
+ "act_iou_n_pairs": 11968,
41
+ "input_r2": 0.9958219333506813,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.7773666875675672,
44
+ "0.7": 0.808761411130152,
45
+ "0.8": 0.8605450812593685,
46
+ "0.9": 0.9230605158056886
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.7774558863196908,
50
+ "0.7": 0.8087822044593573,
51
+ "0.8": 0.8605628646358525,
52
+ "0.9": 0.9230470936609608
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.6617072958728386,
58
+ "mcc_abs": 0.661709934566581,
59
+ "act_iou": 0.43193356294551827,
60
+ "act_iou_n_pairs": 11950,
61
+ "input_r2": 0.9957864080737974,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.7743368405954638,
64
+ "0.7": 0.8092552006838234,
65
+ "0.8": 0.8589086977577072,
66
+ "0.9": 0.9239851096283606
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.7739495397066156,
70
+ "0.7": 0.8092424706708758,
71
+ "0.8": 0.8589469247971228,
72
+ "0.9": 0.9239356466929539
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.6614051608671628,
78
+ "mcc_abs": 0.6614087351942567,
79
+ "act_iou": 0.43308057597145,
80
+ "act_iou_n_pairs": 11934,
81
+ "input_r2": 0.9958052945449676,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.7776594931486868,
84
+ "0.7": 0.8116377053185859,
85
+ "0.8": 0.8588067214655751,
86
+ "0.9": 0.921725572265176
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.7778359861897952,
90
+ "0.7": 0.8116536006278515,
91
+ "0.8": 0.858804064854087,
92
+ "0.9": 0.9217434735942291
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 35.23610981409425,
106
+ "l0_std": 0.141079941279443,
107
+ "recon_r2_mean": 0.9498486990361091,
108
+ "recon_r2_std": 5.304459211779689e-06,
109
+ "l1_per_token_mean": 50.48430344297702,
110
+ "l1_per_token_std": 0.03271937684463371,
111
+ "input_r2_mean": 0.9958045453231487,
112
+ "input_r2_std": 1.7774485208285367e-05,
113
+ "mcc_signed_mean": 0.6622846436961356,
114
+ "mcc_signed_std": 0.001270664413784403,
115
+ "mcc_abs_mean": 0.6622874126297457,
116
+ "mcc_abs_std": 0.001270028303883967,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.7764543404372392,
119
+ "0.7": 0.8098847723775204,
120
+ "0.8": 0.8594201668275502,
121
+ "0.9": 0.9229237325664084
122
+ },
123
+ "matched_mcc_signed_band": 0.8421707530521795,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.776413804072034,
126
+ "0.7": 0.8098927585860283,
127
+ "0.8": 0.8594379514290208,
128
+ "0.9": 0.9229087379827146
129
+ },
130
+ "matched_mcc_abs_band": 0.8421633130174494,
131
+ "act_iou_mean": 0.4324115009318859,
132
+ "act_iou_std": 0.0005969168704624915,
133
+ "act_iou_n_pairs": 11950.666666666666,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": 36,
22
+ "l0_raw": 35.6352,
23
+ "recon_r2": 0.9498,
24
+ "l1_per_token": 50.5619,
25
+ "stats_source": "stats",
26
+ "n_dead": 663,
27
+ "dead_rate": 0.053955078125,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 7747,
34
+ "0.6": 0,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 3136,
41
+ "0.6": 0,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 7,
48
+ "0.6": 1,
49
+ "0.7": 1,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.4743391275405884,
54
+ "n_dup_abs": {
55
+ "0.5": 7739,
56
+ "0.6": 8,
57
+ "0.7": 6,
58
+ "0.8": 0,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 3135,
63
+ "0.6": 4,
64
+ "0.7": 3,
65
+ "0.8": 0,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 7,
70
+ "0.6": 2,
71
+ "0.7": 2,
72
+ "0.8": 1,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.47505998611450195
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 36,
81
+ "l0_raw": 35.6352,
82
+ "recon_r2": 0.9498,
83
+ "l1_per_token": 50.5098,
84
+ "stats_source": "stats",
85
+ "n_dead": 731,
86
+ "dead_rate": 0.059488932291666664,
87
+ "alive_source": "stats",
88
+ "n_dense": 0,
89
+ "dense_rate": 0.0,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 7670,
93
+ "0.6": 0,
94
+ "0.7": 0,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 3106,
100
+ "0.6": 0,
101
+ "0.7": 0,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 8,
107
+ "0.6": 1,
108
+ "0.7": 1,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.47271767258644104,
113
+ "n_dup_abs": {
114
+ "0.5": 7670,
115
+ "0.6": 8,
116
+ "0.7": 4,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 3107,
122
+ "0.6": 4,
123
+ "0.7": 2,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 8,
129
+ "0.6": 2,
130
+ "0.7": 2,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.47340402007102966
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 36,
140
+ "l0_raw": 35.6352,
141
+ "recon_r2": 0.9498,
142
+ "l1_per_token": 50.4943,
143
+ "stats_source": "stats",
144
+ "n_dead": 717,
145
+ "dead_rate": 0.058349609375,
146
+ "alive_source": "stats",
147
+ "n_dense": 0,
148
+ "dense_rate": 0.0,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 7624,
152
+ "0.6": 0,
153
+ "0.7": 0,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 3085,
159
+ "0.6": 0,
160
+ "0.7": 0,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 8,
166
+ "0.6": 1,
167
+ "0.7": 1,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.47238266468048096,
172
+ "n_dup_abs": {
173
+ "0.5": 7630,
174
+ "0.6": 10,
175
+ "0.7": 2,
176
+ "0.8": 0,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 3088,
181
+ "0.6": 5,
182
+ "0.7": 1,
183
+ "0.8": 0,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 8,
188
+ "0.6": 2,
189
+ "0.7": 2,
190
+ "0.8": 1,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.47302475571632385
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.7470758557319641,
199
+ "dcs_abs": 0.7475762367248535,
200
+ "frac_sign_flipped": 0.036865234375,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 11625,
204
+ "n_alive_b": 11557,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.7670451998710632,
207
+ "0.6": 0.7866122722625732,
208
+ "0.7": 0.7866122722625732,
209
+ "0.8": 0.7866122722625732,
210
+ "0.9": 0.7866122722625732
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.5699462890625,
214
+ "0.6": 0.9432779947916666,
215
+ "0.7": 0.9432779947916666,
216
+ "0.8": 0.9432779947916666,
217
+ "0.9": 0.9432779947916666
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.4371745652097161,
221
+ "0.6": 0.7419940468583566,
222
+ "0.7": 0.7419940468583566,
223
+ "0.8": 0.7419940468583566,
224
+ "0.9": 0.7419940468583566
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.4957682291666667,
228
+ "0.6": 0.8229166666666666,
229
+ "0.7": 0.739013671875,
230
+ "0.8": 0.5675455729166666,
231
+ "0.9": 0.31640625
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.7676906585693359,
235
+ "0.6": 0.7868853807449341,
236
+ "0.7": 0.7868661284446716,
237
+ "0.8": 0.7868930697441101,
238
+ "0.9": 0.7868930697441101
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.5702718098958334,
242
+ "0.6": 0.9429524739583334,
243
+ "0.7": 0.9430745442708334,
244
+ "0.8": 0.9432779947916666,
245
+ "0.9": 0.9432779947916666
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.4377923413024595,
249
+ "0.6": 0.7419955164950807,
250
+ "0.7": 0.7420734154851137,
251
+ "0.8": 0.7422589169436833,
252
+ "0.9": 0.7422589169436833
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.4954427083333333,
256
+ "0.6": 0.822265625,
257
+ "0.7": 0.7392578125,
258
+ "0.8": 0.567626953125,
259
+ "0.9": 0.3164876302083333
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.7460986971855164,
264
+ "dcs_abs": 0.7465710639953613,
265
+ "frac_sign_flipped": 0.03564453125,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 11625,
269
+ "n_alive_b": 11571,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.7655117511749268,
272
+ "0.6": 0.7846922278404236,
273
+ "0.7": 0.7846922278404236,
274
+ "0.8": 0.7846922278404236,
275
+ "0.9": 0.7846922278404236
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.571533203125,
279
+ "0.6": 0.94384765625,
280
+ "0.7": 0.94384765625,
281
+ "0.8": 0.94384765625,
282
+ "0.9": 0.94384765625
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.43751538317883387,
286
+ "0.6": 0.7406299201247748,
287
+ "0.7": 0.7406299201247748,
288
+ "0.8": 0.7406299201247748,
289
+ "0.9": 0.7406299201247748
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.4976399739583333,
293
+ "0.6": 0.8221028645833334,
294
+ "0.7": 0.7334798177083334,
295
+ "0.8": 0.5707194010416666,
296
+ "0.9": 0.3097330729166667
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.765274703502655,
300
+ "0.6": 0.7849535942077637,
301
+ "0.7": 0.7848325371742249,
302
+ "0.8": 0.7849403619766235,
303
+ "0.9": 0.7849403619766235
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.5716959635416666,
307
+ "0.6": 0.9434814453125,
308
+ "0.7": 0.9436848958333334,
309
+ "0.8": 0.94384765625,
310
+ "0.9": 0.94384765625
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.4375044589930136,
314
+ "0.6": 0.7405891515663825,
315
+ "0.7": 0.7406346110898692,
316
+ "0.8": 0.7408641209476627,
317
+ "0.9": 0.7408641209476627
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.4977213541666667,
321
+ "0.6": 0.82275390625,
322
+ "0.7": 0.7335611979166666,
323
+ "0.8": 0.5709635416666666,
324
+ "0.9": 0.3099772135416667
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.7451217174530029,
329
+ "dcs_abs": 0.7456211447715759,
330
+ "frac_sign_flipped": 0.037679036458333336,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 11557,
334
+ "n_alive_b": 11571,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.765704870223999,
337
+ "0.6": 0.7844005823135376,
338
+ "0.7": 0.7844005823135376,
339
+ "0.8": 0.7844005823135376,
340
+ "0.9": 0.7844005823135376
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.5706787109375,
344
+ "0.6": 0.9410807291666666,
345
+ "0.7": 0.9410807291666666,
346
+ "0.8": 0.9410807291666666,
347
+ "0.9": 0.9410807291666666
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.4369714682979975,
351
+ "0.6": 0.7381842719623819,
352
+ "0.7": 0.7381842719623819,
353
+ "0.8": 0.7381842719623819,
354
+ "0.9": 0.7381842719623819
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.4961751302083333,
358
+ "0.6": 0.8173828125,
359
+ "0.7": 0.730712890625,
360
+ "0.8": 0.5730794270833334,
361
+ "0.9": 0.3173014322916667
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.7659049034118652,
365
+ "0.6": 0.7845613360404968,
366
+ "0.7": 0.7846297025680542,
367
+ "0.8": 0.7846192717552185,
368
+ "0.9": 0.7846192717552185
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.5705973307291666,
372
+ "0.6": 0.9407145182291666,
373
+ "0.7": 0.9409586588541666,
374
+ "0.8": 0.9410807291666666,
375
+ "0.9": 0.9410807291666666
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.4370232934791905,
379
+ "0.6": 0.7380482392545673,
380
+ "0.7": 0.7383041126255799,
381
+ "0.8": 0.73839007638162,
382
+ "0.9": 0.73839007638162
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.4960123697916667,
386
+ "0.6": 0.8167317708333334,
387
+ "0.7": 0.7303873697916666,
388
+ "0.8": 0.5733235677083334,
389
+ "0.9": 0.3172200520833333
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 704,
403
+ "n_dead_std": 35.90728802532062,
404
+ "n_dense_mean": 0,
405
+ "n_dense_std": 0.0,
406
+ "l0_mean": 36,
407
+ "l0_std": 0.0,
408
+ "recon_r2_mean": 0.9498000000000001,
409
+ "recon_r2_std": 1.3597399555105182e-16,
410
+ "l1_per_token_mean": 50.522,
411
+ "l1_per_token_std": 0.03541285077482469,
412
+ "babel_mu1_signed_mean": 0.4731464882691701,
413
+ "babel_mu1_signed_std": 0.001046350275032571,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 7680.333333333333,
416
+ "0.6": 0.0,
417
+ "0.7": 0.0,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 0.0,
422
+ "n_dup_signed_band_std": 0.0,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 3109.0,
425
+ "0.6": 0.0,
426
+ "0.7": 0.0,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 0.0,
431
+ "n_dup_clusters_signed_band_std": 0.0,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 7.666666666666667,
434
+ "0.6": 1.0,
435
+ "0.7": 1.0,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 1.0,
440
+ "largest_dup_signed_band_std": 0.0,
441
+ "dcs_signed_mean": 0.7460987567901611,
442
+ "dcs_signed_std": 0.0009770691408441255,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.43722047222884913,
445
+ "0.6": 0.7402694129818377,
446
+ "0.7": 0.7402694129818377,
447
+ "0.8": 0.7402694129818377,
448
+ "0.9": 0.7402694129818377
449
+ },
450
+ "dedup_dcs_signed_band": 0.7402694129818377,
451
+ "dedup_dcs_signed_band_std": 0.0019303031470681428,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.766087273756663,
454
+ "0.6": 0.7852350274721781,
455
+ "0.7": 0.7852350274721781,
456
+ "0.8": 0.7852350274721781,
457
+ "0.9": 0.7852350274721781
458
+ },
459
+ "matched_cos_signed_band": 0.7852350274721781,
460
+ "matched_cos_signed_band_std": 0.0012016100398386264,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.5707194010416666,
463
+ "0.6": 0.9427354600694443,
464
+ "0.7": 0.9427354600694443,
465
+ "0.8": 0.9427354600694443,
466
+ "0.9": 0.9427354600694443
467
+ },
468
+ "unique_frac_signed_band": 0.9427354600694443,
469
+ "unique_frac_signed_band_std": 0.0014610712900928126,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.49652777777777773,
472
+ "0.6": 0.82080078125,
473
+ "0.7": 0.7344021267361112,
474
+ "0.8": 0.5704481336805555,
475
+ "0.9": 0.31448025173611116
476
+ },
477
+ "matched_frac_signed_band": 0.6100328233506944,
478
+ "matched_frac_signed_band_std": 0.0012819535547283407,
479
+ "babel_mu1_abs_mean": 0.47382958730061847,
480
+ "babel_mu1_abs_std": 0.0010822990777007893,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 7679.666666666667,
483
+ "0.6": 8.666666666666666,
484
+ "0.7": 4.0,
485
+ "0.8": 0.0,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 3.1666666666666665,
489
+ "n_dup_abs_band_std": 0.2886751345948129,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 3110.0,
492
+ "0.6": 4.333333333333333,
493
+ "0.7": 2.0,
494
+ "0.8": 0.0,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 1.5833333333333333,
498
+ "n_dup_clusters_abs_band_std": 0.14433756729740646,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 7.666666666666667,
501
+ "0.6": 2.0,
502
+ "0.7": 2.0,
503
+ "0.8": 1.0,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 1.5,
507
+ "largest_dup_abs_band_std": 0.0,
508
+ "dcs_abs_mean": 0.7465894818305969,
509
+ "dcs_abs_std": 0.0009776760961247685,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.4374400312582212,
512
+ "0.6": 0.7402109691053435,
513
+ "0.7": 0.7403373797335209,
514
+ "0.8": 0.740504371424322,
515
+ "0.9": 0.740504371424322
516
+ },
517
+ "dedup_dcs_abs_band": 0.7403892729218771,
518
+ "dedup_dcs_abs_band_std": 0.001955247489450024,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.7662900884946188,
521
+ "0.6": 0.7854667703310648,
522
+ "0.7": 0.7854427893956503,
523
+ "0.8": 0.785484234491984,
524
+ "0.9": 0.785484234491984
525
+ },
526
+ "matched_cos_abs_band": 0.7854695071776707,
527
+ "matched_cos_abs_band_std": 0.0012350654052081485,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.5708550347222222,
530
+ "0.6": 0.9423828125,
531
+ "0.7": 0.9425726996527778,
532
+ "0.8": 0.9427354600694443,
533
+ "0.9": 0.9427354600694443
534
+ },
535
+ "unique_frac_abs_band": 0.9426066080729165,
536
+ "unique_frac_abs_band_std": 0.0014553113120434982,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.49639214409722227,
539
+ "0.6": 0.8205837673611112,
540
+ "0.7": 0.734402126736111,
541
+ "0.8": 0.5706380208333334,
542
+ "0.9": 0.3145616319444444
543
+ },
544
+ "matched_frac_abs_band": 0.61004638671875,
545
+ "matched_frac_abs_band_std": 0.0011815904618922564,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 12288
549
+ }
550
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 50,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.8,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 40,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1853,
3
+ "r2_score": 0.9498,
4
+ "l0_per_token_0": 0.0029,
5
+ "dead_rate": 0.054,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 0.9968,
8
+ "l1_per_token": 50.5619
9
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 50,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.8,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 40,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1852,
3
+ "r2_score": 0.9498,
4
+ "l0_per_token_0": 0.0029,
5
+ "dead_rate": 0.0595,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 0.996,
8
+ "l1_per_token": 50.5098
9
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 50,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.8,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 40,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1853,
3
+ "r2_score": 0.9498,
4
+ "l0_per_token_0": 0.0029,
5
+ "dead_rate": 0.0583,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 0.9962,
8
+ "l1_per_token": 50.4943
9
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 71.09119844357977,
20
+ "recon_r2": 0.9523841471073972,
21
+ "l1_per_token": 58.39136403615936
22
+ },
23
+ "seed1": {
24
+ "l0": 71.40945525291829,
25
+ "recon_r2": 0.9523338386197744,
26
+ "l1_per_token": 58.214114896910054
27
+ },
28
+ "seed2": {
29
+ "l0": 71.57131128404669,
30
+ "recon_r2": 0.9523843082278608,
31
+ "l1_per_token": 58.60211755138846
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.7500278227896741,
38
+ "mcc_abs": 0.7500304994777688,
39
+ "act_iou": 0.4551998827895913,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9974972230437296,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.7858687561270252,
44
+ "0.7": 0.8195154540204735,
45
+ "0.8": 0.8685762665006791,
46
+ "0.9": 0.9276617116769719
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.7858070970278039,
50
+ "0.7": 0.8192746433215572,
51
+ "0.8": 0.8683488181168152,
52
+ "0.9": 0.9277006901597833
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.7499596544415255,
58
+ "mcc_abs": 0.7499637282540154,
59
+ "act_iou": 0.4561089470114304,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.9974837164503675,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.7850181415669988,
64
+ "0.7": 0.8194113351448532,
65
+ "0.8": 0.8680070591927237,
66
+ "0.9": 0.9293538753471706
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.7850494097786439,
70
+ "0.7": 0.8192858966549952,
71
+ "0.8": 0.868028628446749,
72
+ "0.9": 0.9293724363777015
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.7511256937414075,
78
+ "mcc_abs": 0.7511288112303944,
79
+ "act_iou": 0.4584235791135581,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.9975041518934918,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.7864068008377323,
84
+ "0.7": 0.8189441775250538,
85
+ "0.8": 0.8682801810551152,
86
+ "0.9": 0.9285019150991416
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.7861995551763159,
90
+ "0.7": 0.8186480005728333,
91
+ "0.8": 0.8682192378841881,
92
+ "0.9": 0.9285198253898055
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 71.35732166018158,
106
+ "l0_std": 0.24426526260201284,
107
+ "recon_r2_mean": 0.9523674313183442,
108
+ "recon_r2_std": 2.9092241884193886e-05,
109
+ "l1_per_token_mean": 58.40253216148596,
110
+ "l1_per_token_std": 0.19424227201630254,
111
+ "input_r2_mean": 0.9974950304625296,
112
+ "input_r2_std": 1.0392660542946744e-05,
113
+ "mcc_signed_mean": 0.7503710569908689,
114
+ "mcc_signed_std": 0.0006544227990143535,
115
+ "mcc_abs_mean": 0.7503743463207262,
116
+ "mcc_abs_std": 0.0006542381630728305,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.7857645661772521,
119
+ "0.7": 0.8192903222301268,
120
+ "0.8": 0.8682878355828394,
121
+ "0.9": 0.9285058340410947
122
+ },
123
+ "matched_mcc_signed_band": 0.8504621395078282,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.7856853539942547,
126
+ "0.7": 0.819069513516462,
127
+ "0.8": 0.8681988948159174,
128
+ "0.9": 0.9285309839757634
129
+ },
130
+ "matched_mcc_abs_band": 0.8503711865755994,
131
+ "act_iou_mean": 0.45657746963819323,
132
+ "act_iou_std": 0.0016621340457769439,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": 71,
22
+ "l0_raw": 71.2704,
23
+ "recon_r2": 0.9524,
24
+ "l1_per_token": 58.4376,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 9445,
34
+ "0.6": 58,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 3534,
41
+ "0.6": 20,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 7,
48
+ "0.6": 4,
49
+ "0.7": 1,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.5134926438331604,
54
+ "n_dup_abs": {
55
+ "0.5": 9439,
56
+ "0.6": 70,
57
+ "0.7": 4,
58
+ "0.8": 0,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 3530,
63
+ "0.6": 26,
64
+ "0.7": 2,
65
+ "0.8": 0,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 7,
70
+ "0.6": 4,
71
+ "0.7": 2,
72
+ "0.8": 1,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.5139548778533936
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 71,
81
+ "l0_raw": 71.2704,
82
+ "recon_r2": 0.9523,
83
+ "l1_per_token": 58.26,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 2,
89
+ "dense_rate": 0.00016276041666666666,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 9388,
93
+ "0.6": 48,
94
+ "0.7": 0,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 3520,
100
+ "0.6": 17,
101
+ "0.7": 0,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 7,
107
+ "0.6": 4,
108
+ "0.7": 1,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.5133753418922424,
113
+ "n_dup_abs": {
114
+ "0.5": 9388,
115
+ "0.6": 66,
116
+ "0.7": 0,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 3521,
122
+ "0.6": 26,
123
+ "0.7": 0,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 7,
129
+ "0.6": 4,
130
+ "0.7": 1,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.513782799243927
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 71,
140
+ "l0_raw": 71.2704,
141
+ "recon_r2": 0.9524,
142
+ "l1_per_token": 58.6484,
143
+ "stats_source": "stats",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 1,
148
+ "dense_rate": 8.138020833333333e-05,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 9350,
152
+ "0.6": 57,
153
+ "0.7": 0,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 3511,
159
+ "0.6": 18,
160
+ "0.7": 0,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 8,
166
+ "0.6": 4,
167
+ "0.7": 1,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.5129317045211792,
172
+ "n_dup_abs": {
173
+ "0.5": 9352,
174
+ "0.6": 73,
175
+ "0.7": 2,
176
+ "0.8": 0,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 3513,
181
+ "0.6": 26,
182
+ "0.7": 1,
183
+ "0.8": 0,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 8,
188
+ "0.6": 4,
189
+ "0.7": 2,
190
+ "0.8": 1,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.5133534669876099
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.7969434857368469,
199
+ "dcs_abs": 0.7970011830329895,
200
+ "frac_sign_flipped": 0.0026041666666666665,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 12288,
204
+ "n_alive_b": 12288,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.7957947850227356,
207
+ "0.6": 0.7960567474365234,
208
+ "0.7": 0.7969434857368469,
209
+ "0.8": 0.7969434857368469,
210
+ "0.9": 0.7969434857368469
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.5207112630208334,
214
+ "0.6": 0.9971923828125,
215
+ "0.7": 1.0,
216
+ "0.8": 1.0,
217
+ "0.9": 1.0
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.41437930761458125,
221
+ "0.6": 0.7938217248301953,
222
+ "0.7": 0.7969434857368469,
223
+ "0.8": 0.7969434857368469,
224
+ "0.9": 0.7969434857368469
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.4825032552083333,
228
+ "0.6": 0.9029134114583334,
229
+ "0.7": 0.790771484375,
230
+ "0.8": 0.5743001302083334,
231
+ "0.9": 0.2783203125
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.7955784797668457,
235
+ "0.6": 0.7960397601127625,
236
+ "0.7": 0.7970454096794128,
237
+ "0.8": 0.7970011234283447,
238
+ "0.9": 0.7970011234283447
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.5208333333333334,
242
+ "0.6": 0.99658203125,
243
+ "0.7": 0.9999186197916666,
244
+ "0.8": 1.0,
245
+ "0.9": 1.0
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.4143637915452322,
249
+ "0.6": 0.7933189210889395,
250
+ "0.7": 0.796980545957922,
251
+ "0.8": 0.7970011234283447,
252
+ "0.9": 0.7970011234283447
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.4833984375,
256
+ "0.6": 0.9022623697916666,
257
+ "0.7": 0.791015625,
258
+ "0.8": 0.57470703125,
259
+ "0.9": 0.2782389322916667
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.7959124445915222,
264
+ "dcs_abs": 0.7959718704223633,
265
+ "frac_sign_flipped": 0.0026041666666666665,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 12288,
269
+ "n_alive_b": 12288,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.7941434979438782,
272
+ "0.6": 0.7948445677757263,
273
+ "0.7": 0.7959124445915222,
274
+ "0.8": 0.7959124445915222,
275
+ "0.9": 0.7959124445915222
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.5218912760416666,
279
+ "0.6": 0.9968668619791666,
280
+ "0.7": 1.0,
281
+ "0.8": 1.0,
282
+ "0.9": 1.0
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.41445656350212323,
286
+ "0.6": 0.7923542100397754,
287
+ "0.7": 0.7959124445915222,
288
+ "0.8": 0.7959124445915222,
289
+ "0.9": 0.7959124445915222
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.4825846354166667,
293
+ "0.6": 0.9031575520833334,
294
+ "0.7": 0.7906087239583334,
295
+ "0.8": 0.57275390625,
296
+ "0.9": 0.2725423177083333
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.7940723896026611,
300
+ "0.6": 0.7949684858322144,
301
+ "0.7": 0.7959861755371094,
302
+ "0.8": 0.7959718704223633,
303
+ "0.9": 0.7959718704223633
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.52197265625,
307
+ "0.6": 0.9962972005208334,
308
+ "0.7": 0.9998779296875,
309
+ "0.8": 1.0,
310
+ "0.9": 1.0
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.4144840744556859,
314
+ "0.6": 0.7920248769369209,
315
+ "0.7": 0.7958890092559159,
316
+ "0.8": 0.7959718704223633,
317
+ "0.9": 0.7959718704223633
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.482177734375,
321
+ "0.6": 0.9024251302083334,
322
+ "0.7": 0.79052734375,
323
+ "0.8": 0.5726725260416666,
324
+ "0.9": 0.2724609375
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.7979528903961182,
329
+ "dcs_abs": 0.7979879379272461,
330
+ "frac_sign_flipped": 0.0021158854166666665,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 12288,
334
+ "n_alive_b": 12288,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.7935064435005188,
337
+ "0.6": 0.7969188690185547,
338
+ "0.7": 0.7979529500007629,
339
+ "0.8": 0.7979529500007629,
340
+ "0.9": 0.7979529500007629
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.5236409505208334,
344
+ "0.6": 0.9971516927083334,
345
+ "0.7": 1.0,
346
+ "0.8": 1.0,
347
+ "0.9": 1.0
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.4155124683190176,
351
+ "0.6": 0.7946489991930624,
352
+ "0.7": 0.7979529500007629,
353
+ "0.8": 0.7979529500007629,
354
+ "0.9": 0.7979529500007629
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.4847005208333333,
358
+ "0.6": 0.9025065104166666,
359
+ "0.7": 0.7923177083333334,
360
+ "0.8": 0.5812174479166666,
361
+ "0.9": 0.27783203125
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.7933503985404968,
365
+ "0.6": 0.7969396710395813,
366
+ "0.7": 0.7980155348777771,
367
+ "0.8": 0.7979879379272461,
368
+ "0.9": 0.7979879379272461
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.523681640625,
372
+ "0.6": 0.9964599609375,
373
+ "0.7": 0.9999593098958334,
374
+ "0.8": 1.0,
375
+ "0.9": 1.0
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.415463038298185,
379
+ "0.6": 0.7941184734736453,
380
+ "0.7": 0.7979830635425363,
381
+ "0.8": 0.7979879379272461,
382
+ "0.9": 0.7979879379272461
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.48486328125,
386
+ "0.6": 0.902587890625,
387
+ "0.7": 0.7928059895833334,
388
+ "0.8": 0.5809733072916666,
389
+ "0.9": 0.2777506510416667
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 0,
403
+ "n_dead_std": 0.0,
404
+ "n_dense_mean": 1,
405
+ "n_dense_std": 1.0,
406
+ "l0_mean": 71,
407
+ "l0_std": 0.0,
408
+ "recon_r2_mean": 0.9523666666666667,
409
+ "recon_r2_std": 5.773502691895622e-05,
410
+ "l1_per_token_mean": 58.44866666666667,
411
+ "l1_per_token_std": 0.19443634776793697,
412
+ "babel_mu1_signed_mean": 0.5132665634155273,
413
+ "babel_mu1_signed_std": 0.00029586786861507457,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 9394.333333333334,
416
+ "0.6": 54.333333333333336,
417
+ "0.7": 0.0,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 13.583333333333334,
422
+ "n_dup_signed_band_std": 1.3768926368215255,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 3521.6666666666665,
425
+ "0.6": 18.333333333333332,
426
+ "0.7": 0.0,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 4.583333333333333,
431
+ "n_dup_clusters_signed_band_std": 0.3818813079129867,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 7.333333333333333,
434
+ "0.6": 4.0,
435
+ "0.7": 1.0,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 1.75,
440
+ "largest_dup_signed_band_std": 0.0,
441
+ "dcs_signed_mean": 0.7969362735748291,
442
+ "dcs_signed_std": 0.0010202420212057672,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.4147827798119074,
445
+ "0.6": 0.7936083113543444,
446
+ "0.7": 0.7969362934430441,
447
+ "0.8": 0.7969362934430441,
448
+ "0.9": 0.7969362934430441
449
+ },
450
+ "dedup_dcs_signed_band": 0.7961042979208692,
451
+ "dedup_dcs_signed_band_std": 0.0010532676659331863,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.7944815754890442,
454
+ "0.6": 0.7959400614102682,
455
+ "0.7": 0.7969362934430441,
456
+ "0.8": 0.7969362934430441,
457
+ "0.9": 0.7969362934430441
458
+ },
459
+ "matched_cos_signed_band": 0.7966872354348501,
460
+ "matched_cos_signed_band_std": 0.0010249144316835713,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.5220811631944445,
463
+ "0.6": 0.9970703125,
464
+ "0.7": 1.0,
465
+ "0.8": 1.0,
466
+ "0.9": 1.0
467
+ },
468
+ "unique_frac_signed_band": 0.999267578125,
469
+ "unique_frac_signed_band_std": 4.434101301619054e-05,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.4832628038194444,
472
+ "0.6": 0.9028591579861112,
473
+ "0.7": 0.791232638888889,
474
+ "0.8": 0.5760904947916666,
475
+ "0.9": 0.2762315538194444
476
+ },
477
+ "matched_frac_signed_band": 0.6366034613715277,
478
+ "matched_frac_signed_band_std": 0.001851548781584755,
479
+ "babel_mu1_abs_mean": 0.5136970480283102,
480
+ "babel_mu1_abs_std": 0.00030973974657040626,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 9393.0,
483
+ "0.6": 69.66666666666667,
484
+ "0.7": 2.0,
485
+ "0.8": 0.0,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 17.916666666666668,
489
+ "n_dup_abs_band_std": 1.2332207155790618,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 3521.3333333333335,
492
+ "0.6": 26.0,
493
+ "0.7": 1.0,
494
+ "0.8": 0.0,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 6.75,
498
+ "n_dup_clusters_abs_band_std": 0.25,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 7.333333333333333,
501
+ "0.6": 4.0,
502
+ "0.7": 1.6666666666666667,
503
+ "0.8": 1.0,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 1.9166666666666667,
507
+ "largest_dup_abs_band_std": 0.14433756729740646,
508
+ "dcs_abs_mean": 0.796986997127533,
509
+ "dcs_abs_std": 0.0010081086131941929,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.4147703014330344,
512
+ "0.6": 0.7931540904998352,
513
+ "0.7": 0.7969508729187914,
514
+ "0.8": 0.796986977259318,
515
+ "0.9": 0.796986977259318
516
+ },
517
+ "dedup_dcs_abs_band": 0.7960197294843157,
518
+ "dedup_dcs_abs_band_std": 0.0010286048900542725,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.7943337559700012,
521
+ "0.6": 0.7959826389948527,
522
+ "0.7": 0.7970157066980997,
523
+ "0.8": 0.796986977259318,
524
+ "0.9": 0.796986977259318
525
+ },
526
+ "matched_cos_abs_band": 0.7967430750528971,
527
+ "matched_cos_abs_band_std": 0.0010043942224412575,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.5221625434027778,
530
+ "0.6": 0.9964463975694445,
531
+ "0.7": 0.9999186197916666,
532
+ "0.8": 1.0,
533
+ "0.9": 1.0
534
+ },
535
+ "unique_frac_abs_band": 0.9990912543402778,
536
+ "unique_frac_abs_band_std": 4.235160317921298e-05,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.4834798177083333,
539
+ "0.6": 0.9024251302083334,
540
+ "0.7": 0.7914496527777778,
541
+ "0.8": 0.5761176215277777,
542
+ "0.9": 0.27615017361111116
543
+ },
544
+ "matched_frac_abs_band": 0.63653564453125,
545
+ "matched_frac_abs_band_std": 0.0020040650844935504,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 12288
549
+ }
550
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 22,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1759,
3
+ "r2_score": 0.9524,
4
+ "l0_per_token_0": 0.0058,
5
+ "l0_per_token_0.001": 0.0058,
6
+ "l0_per_token_0.01": 0.0058,
7
+ "l0_per_token_0.05": 0.0058,
8
+ "l0_per_token_0.1": 0.0057,
9
+ "l0_per_token_0.5": 0.0029,
10
+ "l0_per_token_1.0": 0.0014,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 58.4376
15
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 22,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1761,
3
+ "r2_score": 0.9523,
4
+ "l0_per_token_0": 0.0058,
5
+ "l0_per_token_0.001": 0.0058,
6
+ "l0_per_token_0.01": 0.0058,
7
+ "l0_per_token_0.05": 0.0058,
8
+ "l0_per_token_0.1": 0.0058,
9
+ "l0_per_token_0.5": 0.0029,
10
+ "l0_per_token_1.0": 0.0013,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 58.26
15
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 22,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1759,
3
+ "r2_score": 0.9524,
4
+ "l0_per_token_0": 0.0058,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 58.6484
9
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 137.0576718547341,
20
+ "recon_r2": 0.9819069950791086,
21
+ "l1_per_token": 149.93086001487765
22
+ },
23
+ "seed1": {
24
+ "l0": 137.27400778210117,
25
+ "recon_r2": 0.9819008258322424,
26
+ "l1_per_token": 149.56244017092325
27
+ },
28
+ "seed2": {
29
+ "l0": 138.21439040207522,
30
+ "recon_r2": 0.9819432031869295,
31
+ "l1_per_token": 151.22428314247435
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.844550231318728,
38
+ "mcc_abs": 0.8445640060318582,
39
+ "act_iou": 0.5074729118922311,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.9924763461473485,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.9306475890032538,
44
+ "0.7": 0.9397792699507539,
45
+ "0.8": 0.9481302892895763,
46
+ "0.9": 0.9557333663229426
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.9294119775515188,
50
+ "0.7": 0.9398317641790239,
51
+ "0.8": 0.9478422470060124,
52
+ "0.9": 0.9557333663395272
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.8429910657341265,
58
+ "mcc_abs": 0.8429911402419448,
59
+ "act_iou": 0.5052272174171811,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9924368617718973,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.9276478806301962,
64
+ "0.7": 0.9368093737129444,
65
+ "0.8": 0.9462548443079576,
66
+ "0.9": 0.9551839539310047
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.9270726312662237,
70
+ "0.7": 0.9366073203611118,
71
+ "0.8": 0.945870991911303,
72
+ "0.9": 0.955183953905391
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.8453560632347036,
78
+ "mcc_abs": 0.8453764790182281,
79
+ "act_iou": 0.5111202678449333,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.9925993963285449,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.9276392516954068,
84
+ "0.7": 0.9386938573077988,
85
+ "0.8": 0.9458195925427115,
86
+ "0.9": 0.9557235413870429
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.927566637924043,
90
+ "0.7": 0.9387584654480161,
91
+ "0.8": 0.9456582618677191,
92
+ "0.9": 0.9557235414261833
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 137.51535667963682,
106
+ "l0_std": 0.6149686310674686,
107
+ "recon_r2_mean": 0.98191700803276,
108
+ "recon_r2_std": 2.2894420241430398e-05,
109
+ "l1_per_token_mean": 150.23919444275842,
110
+ "l1_per_token_std": 0.8727732266092169,
111
+ "input_r2_mean": 0.9925042014159301,
112
+ "input_r2_std": 8.477209161616265e-05,
113
+ "mcc_signed_mean": 0.8442991200958527,
114
+ "mcc_signed_std": 0.0012023293763043362,
115
+ "mcc_abs_mean": 0.8443105417640103,
116
+ "mcc_abs_std": 0.001212700857885798,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.9286449071096189,
119
+ "0.7": 0.9384275003238324,
120
+ "0.8": 0.9467349087134153,
121
+ "0.9": 0.95554695388033
122
+ },
123
+ "matched_mcc_signed_band": 0.9423385675067991,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.9280170822472619,
126
+ "0.7": 0.938399183329384,
127
+ "0.8": 0.9464571669283447,
128
+ "0.9": 0.9555469538903671
129
+ },
130
+ "matched_mcc_abs_band": 0.9421050965988395,
131
+ "act_iou_mean": 0.5079401323847819,
132
+ "act_iou_std": 0.0029741775463986474,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/dict_metrics_step_final.json ADDED
@@ -0,0 +1,595 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 3,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 3072,
20
+ "d_in": 768,
21
+ "l0": 138,
22
+ "l0_raw": 137.6256,
23
+ "recon_r2": 0.9819,
24
+ "l1_per_token": 150.1596,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 85,
30
+ "dense_rate": 0.027669270833333332,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 7,
34
+ "0.6": 0,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 3,
41
+ "0.6": 0,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "n_dup_removed_signed": {
47
+ "0.5": 4,
48
+ "0.6": 0,
49
+ "0.7": 0,
50
+ "0.8": 0,
51
+ "0.9": 0
52
+ },
53
+ "largest_dup_signed": {
54
+ "0.5": 3,
55
+ "0.6": 1,
56
+ "0.7": 1,
57
+ "0.8": 1,
58
+ "0.9": 1
59
+ },
60
+ "babel_mu1_signed": 0.18895398080348969,
61
+ "n_dup_abs": {
62
+ "0.5": 49,
63
+ "0.6": 28,
64
+ "0.7": 12,
65
+ "0.8": 4,
66
+ "0.9": 2
67
+ },
68
+ "n_dup_clusters_abs": {
69
+ "0.5": 23,
70
+ "0.6": 14,
71
+ "0.7": 6,
72
+ "0.8": 2,
73
+ "0.9": 1
74
+ },
75
+ "n_dup_removed_abs": {
76
+ "0.5": 26,
77
+ "0.6": 14,
78
+ "0.7": 6,
79
+ "0.8": 2,
80
+ "0.9": 1
81
+ },
82
+ "largest_dup_abs": {
83
+ "0.5": 4,
84
+ "0.6": 2,
85
+ "0.7": 2,
86
+ "0.8": 2,
87
+ "0.9": 2
88
+ },
89
+ "babel_mu1_abs": 0.2072591334581375,
90
+ "dup_excludes_dead": true
91
+ },
92
+ "seed1": {
93
+ "d_sae": 3072,
94
+ "d_in": 768,
95
+ "l0": 138,
96
+ "l0_raw": 137.93280000000001,
97
+ "recon_r2": 0.9819,
98
+ "l1_per_token": 149.8765,
99
+ "stats_source": "stats",
100
+ "n_dead": 0,
101
+ "dead_rate": 0.0,
102
+ "alive_source": "stats",
103
+ "n_dense": 91,
104
+ "dense_rate": 0.029622395833333332,
105
+ "dense_threshold": 0.1,
106
+ "n_dup_signed": {
107
+ "0.5": 9,
108
+ "0.6": 0,
109
+ "0.7": 0,
110
+ "0.8": 0,
111
+ "0.9": 0
112
+ },
113
+ "n_dup_clusters_signed": {
114
+ "0.5": 4,
115
+ "0.6": 0,
116
+ "0.7": 0,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_removed_signed": {
121
+ "0.5": 5,
122
+ "0.6": 0,
123
+ "0.7": 0,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_signed": {
128
+ "0.5": 3,
129
+ "0.6": 1,
130
+ "0.7": 1,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_signed": 0.18931181728839874,
135
+ "n_dup_abs": {
136
+ "0.5": 49,
137
+ "0.6": 26,
138
+ "0.7": 16,
139
+ "0.8": 2,
140
+ "0.9": 2
141
+ },
142
+ "n_dup_clusters_abs": {
143
+ "0.5": 23,
144
+ "0.6": 13,
145
+ "0.7": 8,
146
+ "0.8": 1,
147
+ "0.9": 1
148
+ },
149
+ "n_dup_removed_abs": {
150
+ "0.5": 26,
151
+ "0.6": 13,
152
+ "0.7": 8,
153
+ "0.8": 1,
154
+ "0.9": 1
155
+ },
156
+ "largest_dup_abs": {
157
+ "0.5": 4,
158
+ "0.6": 2,
159
+ "0.7": 2,
160
+ "0.8": 2,
161
+ "0.9": 2
162
+ },
163
+ "babel_mu1_abs": 0.20621319115161896,
164
+ "dup_excludes_dead": true
165
+ },
166
+ "seed2": {
167
+ "d_sae": 3072,
168
+ "d_in": 768,
169
+ "l0": 139,
170
+ "l0_raw": 138.5472,
171
+ "recon_r2": 0.9819,
172
+ "l1_per_token": 151.3502,
173
+ "stats_source": "stats",
174
+ "n_dead": 0,
175
+ "dead_rate": 0.0,
176
+ "alive_source": "stats",
177
+ "n_dense": 86,
178
+ "dense_rate": 0.027994791666666668,
179
+ "dense_threshold": 0.1,
180
+ "n_dup_signed": {
181
+ "0.5": 12,
182
+ "0.6": 0,
183
+ "0.7": 0,
184
+ "0.8": 0,
185
+ "0.9": 0
186
+ },
187
+ "n_dup_clusters_signed": {
188
+ "0.5": 5,
189
+ "0.6": 0,
190
+ "0.7": 0,
191
+ "0.8": 0,
192
+ "0.9": 0
193
+ },
194
+ "n_dup_removed_signed": {
195
+ "0.5": 7,
196
+ "0.6": 0,
197
+ "0.7": 0,
198
+ "0.8": 0,
199
+ "0.9": 0
200
+ },
201
+ "largest_dup_signed": {
202
+ "0.5": 4,
203
+ "0.6": 1,
204
+ "0.7": 1,
205
+ "0.8": 1,
206
+ "0.9": 1
207
+ },
208
+ "babel_mu1_signed": 0.18924765288829803,
209
+ "n_dup_abs": {
210
+ "0.5": 50,
211
+ "0.6": 18,
212
+ "0.7": 8,
213
+ "0.8": 4,
214
+ "0.9": 0
215
+ },
216
+ "n_dup_clusters_abs": {
217
+ "0.5": 23,
218
+ "0.6": 9,
219
+ "0.7": 4,
220
+ "0.8": 2,
221
+ "0.9": 0
222
+ },
223
+ "n_dup_removed_abs": {
224
+ "0.5": 27,
225
+ "0.6": 9,
226
+ "0.7": 4,
227
+ "0.8": 2,
228
+ "0.9": 0
229
+ },
230
+ "largest_dup_abs": {
231
+ "0.5": 5,
232
+ "0.6": 2,
233
+ "0.7": 2,
234
+ "0.8": 2,
235
+ "0.9": 1
236
+ },
237
+ "babel_mu1_abs": 0.20668457448482513,
238
+ "dup_excludes_dead": true
239
+ }
240
+ },
241
+ "per_pair": {
242
+ "seed0\u2194seed1": {
243
+ "dcs_signed": 0.8818853497505188,
244
+ "dcs_abs": 0.8827230334281921,
245
+ "frac_sign_flipped": 0.014973958333333334,
246
+ "alive_source": "stats",
247
+ "frac_denominator": "d_sae",
248
+ "n_alive_a": 3072,
249
+ "n_alive_b": 3072,
250
+ "matched_cos_signed": {
251
+ "0.5": 0.8822191953659058,
252
+ "0.6": 0.8818853497505188,
253
+ "0.7": 0.8818853497505188,
254
+ "0.8": 0.8818853497505188,
255
+ "0.9": 0.8818853497505188
256
+ },
257
+ "unique_frac_signed": {
258
+ "0.5": 0.99853515625,
259
+ "0.6": 1.0,
260
+ "0.7": 1.0,
261
+ "0.8": 1.0,
262
+ "0.9": 1.0
263
+ },
264
+ "dedup_dcs_signed": {
265
+ "0.5": 0.880926882091444,
266
+ "0.6": 0.8818853497505188,
267
+ "0.7": 0.8818853497505188,
268
+ "0.8": 0.8818853497505188,
269
+ "0.9": 0.8818853497505188
270
+ },
271
+ "matched_frac_signed": {
272
+ "0.5": 0.9000651041666666,
273
+ "0.6": 0.8756510416666666,
274
+ "0.7": 0.8525390625,
275
+ "0.8": 0.826171875,
276
+ "0.9": 0.7867838541666666
277
+ },
278
+ "matched_cos_abs": {
279
+ "0.5": 0.8850637078285217,
280
+ "0.6": 0.8843180537223816,
281
+ "0.7": 0.8836691975593567,
282
+ "0.8": 0.8830690383911133,
283
+ "0.9": 0.8829354047775269
284
+ },
285
+ "unique_frac_abs": {
286
+ "0.5": 0.9915364583333334,
287
+ "0.6": 0.99560546875,
288
+ "0.7": 0.9977213541666666,
289
+ "0.8": 0.99951171875,
290
+ "0.9": 0.9996744791666666
291
+ },
292
+ "dedup_dcs_abs": {
293
+ "0.5": 0.8775729342596605,
294
+ "0.6": 0.8804318904003594,
295
+ "0.7": 0.881655628424293,
296
+ "0.8": 0.8826378523372114,
297
+ "0.9": 0.8826479909087841
298
+ },
299
+ "matched_frac_abs": {
300
+ "0.5": 0.8971354166666666,
301
+ "0.6": 0.8753255208333334,
302
+ "0.7": 0.8518880208333334,
303
+ "0.8": 0.826171875,
304
+ "0.9": 0.7867838541666666
305
+ }
306
+ },
307
+ "seed0\u2194seed2": {
308
+ "dcs_signed": 0.8805902600288391,
309
+ "dcs_abs": 0.8810338973999023,
310
+ "frac_sign_flipped": 0.012369791666666666,
311
+ "alive_source": "stats",
312
+ "frac_denominator": "d_sae",
313
+ "n_alive_a": 3072,
314
+ "n_alive_b": 3072,
315
+ "matched_cos_signed": {
316
+ "0.5": 0.8807594776153564,
317
+ "0.6": 0.8805902600288391,
318
+ "0.7": 0.8805902600288391,
319
+ "0.8": 0.8805902600288391,
320
+ "0.9": 0.8805902600288391
321
+ },
322
+ "unique_frac_signed": {
323
+ "0.5": 0.9982096354166666,
324
+ "0.6": 1.0,
325
+ "0.7": 1.0,
326
+ "0.8": 1.0,
327
+ "0.9": 1.0
328
+ },
329
+ "dedup_dcs_signed": {
330
+ "0.5": 0.8791825970401987,
331
+ "0.6": 0.8805902600288391,
332
+ "0.7": 0.8805902600288391,
333
+ "0.8": 0.8805902600288391,
334
+ "0.9": 0.8805902600288391
335
+ },
336
+ "matched_frac_signed": {
337
+ "0.5": 0.9013671875,
338
+ "0.6": 0.8766276041666666,
339
+ "0.7": 0.8551432291666666,
340
+ "0.8": 0.8209635416666666,
341
+ "0.9": 0.7786458333333334
342
+ },
343
+ "matched_cos_abs": {
344
+ "0.5": 0.8830110430717468,
345
+ "0.6": 0.8825953602790833,
346
+ "0.7": 0.8815736174583435,
347
+ "0.8": 0.8810485601425171,
348
+ "0.9": 0.8812444806098938
349
+ },
350
+ "unique_frac_abs": {
351
+ "0.5": 0.9913736979166666,
352
+ "0.6": 0.9962565104166666,
353
+ "0.7": 0.9983723958333334,
354
+ "0.8": 0.9993489583333334,
355
+ "0.9": 0.9998372395833334
356
+ },
357
+ "dedup_dcs_abs": {
358
+ "0.5": 0.8753939230712906,
359
+ "0.6": 0.8792913737415802,
360
+ "0.7": 0.8801387645653449,
361
+ "0.8": 0.8804749608195077,
362
+ "0.9": 0.8811010488910446
363
+ },
364
+ "matched_frac_abs": {
365
+ "0.5": 0.8974609375,
366
+ "0.6": 0.875,
367
+ "0.7": 0.8544921875,
368
+ "0.8": 0.8206380208333334,
369
+ "0.9": 0.7786458333333334
370
+ }
371
+ },
372
+ "seed1\u2194seed2": {
373
+ "dcs_signed": 0.8839121460914612,
374
+ "dcs_abs": 0.8843576312065125,
375
+ "frac_sign_flipped": 0.012044270833333334,
376
+ "alive_source": "stats",
377
+ "frac_denominator": "d_sae",
378
+ "n_alive_a": 3072,
379
+ "n_alive_b": 3072,
380
+ "matched_cos_signed": {
381
+ "0.5": 0.8841296434402466,
382
+ "0.6": 0.8839120864868164,
383
+ "0.7": 0.8839120864868164,
384
+ "0.8": 0.8839120864868164,
385
+ "0.9": 0.8839120864868164
386
+ },
387
+ "unique_frac_signed": {
388
+ "0.5": 0.998046875,
389
+ "0.6": 1.0,
390
+ "0.7": 1.0,
391
+ "0.8": 1.0,
392
+ "0.9": 1.0
393
+ },
394
+ "dedup_dcs_signed": {
395
+ "0.5": 0.8824028277304024,
396
+ "0.6": 0.8839120864868164,
397
+ "0.7": 0.8839120864868164,
398
+ "0.8": 0.8839120864868164,
399
+ "0.9": 0.8839120864868164
400
+ },
401
+ "matched_frac_signed": {
402
+ "0.5": 0.9000651041666666,
403
+ "0.6": 0.8824869791666666,
404
+ "0.7": 0.857421875,
405
+ "0.8": 0.8326822916666666,
406
+ "0.9": 0.7861328125
407
+ },
408
+ "matched_cos_abs": {
409
+ "0.5": 0.8862475752830505,
410
+ "0.6": 0.8858160972595215,
411
+ "0.7": 0.8855254650115967,
412
+ "0.8": 0.8845116496086121,
413
+ "0.9": 0.8845872282981873
414
+ },
415
+ "unique_frac_abs": {
416
+ "0.5": 0.9913736979166666,
417
+ "0.6": 0.9964192708333334,
418
+ "0.7": 0.998046875,
419
+ "0.8": 0.99951171875,
420
+ "0.9": 0.9998372395833334
421
+ },
422
+ "dedup_dcs_abs": {
423
+ "0.5": 0.8786025359780372,
424
+ "0.6": 0.8826442297237616,
425
+ "0.7": 0.8837959230877459,
426
+ "0.8": 0.8840797591547016,
427
+ "0.9": 0.8844432525123315
428
+ },
429
+ "matched_frac_abs": {
430
+ "0.5": 0.8948567708333334,
431
+ "0.6": 0.8798828125,
432
+ "0.7": 0.8570963541666666,
433
+ "0.8": 0.8323567708333334,
434
+ "0.9": 0.7861328125
435
+ }
436
+ }
437
+ },
438
+ "summary": {
439
+ "n_seeds": 3,
440
+ "n_pairs": 3,
441
+ "report_band": [
442
+ 0.6,
443
+ 0.7,
444
+ 0.8,
445
+ 0.9
446
+ ],
447
+ "n_dead_mean": 0,
448
+ "n_dead_std": 0.0,
449
+ "n_dense_mean": 87,
450
+ "n_dense_std": 3.214550253664318,
451
+ "l0_mean": 138,
452
+ "l0_std": 0.46925575116348217,
453
+ "recon_r2_mean": 0.9819,
454
+ "recon_r2_std": 0.0,
455
+ "l1_per_token_mean": 150.4621,
456
+ "l1_per_token_std": 0.7820342767423953,
457
+ "babel_mu1_signed_mean": 0.18917115032672882,
458
+ "babel_mu1_signed_std": 0.0001907910347160109,
459
+ "n_dup_signed_per_threshold": {
460
+ "0.5": 9.333333333333334,
461
+ "0.6": 0.0,
462
+ "0.7": 0.0,
463
+ "0.8": 0.0,
464
+ "0.9": 0.0
465
+ },
466
+ "n_dup_signed_band": 0.0,
467
+ "n_dup_signed_band_std": 0.0,
468
+ "n_dup_clusters_signed_per_threshold": {
469
+ "0.5": 4.0,
470
+ "0.6": 0.0,
471
+ "0.7": 0.0,
472
+ "0.8": 0.0,
473
+ "0.9": 0.0
474
+ },
475
+ "n_dup_clusters_signed_band": 0.0,
476
+ "n_dup_clusters_signed_band_std": 0.0,
477
+ "largest_dup_signed_per_threshold": {
478
+ "0.5": 3.3333333333333335,
479
+ "0.6": 1.0,
480
+ "0.7": 1.0,
481
+ "0.8": 1.0,
482
+ "0.9": 1.0
483
+ },
484
+ "largest_dup_signed_band": 1.0,
485
+ "largest_dup_signed_band_std": 0.0,
486
+ "dcs_signed_mean": 0.8821292519569397,
487
+ "dcs_signed_std": 0.0016743201509817132,
488
+ "dedup_dcs_signed_per_threshold": {
489
+ "0.5": 0.8808374356206817,
490
+ "0.6": 0.8821292320887247,
491
+ "0.7": 0.8821292320887247,
492
+ "0.8": 0.8821292320887247,
493
+ "0.9": 0.8821292320887247
494
+ },
495
+ "dedup_dcs_signed_band": 0.8821292320887247,
496
+ "dedup_dcs_signed_band_std": 0.001674288416132688,
497
+ "matched_cos_signed_per_threshold": {
498
+ "0.5": 0.8823694388071696,
499
+ "0.6": 0.8821292320887247,
500
+ "0.7": 0.8821292320887247,
501
+ "0.8": 0.8821292320887247,
502
+ "0.9": 0.8821292320887247
503
+ },
504
+ "matched_cos_signed_band": 0.8821292320887247,
505
+ "matched_cos_signed_band_std": 0.001674288416132688,
506
+ "unique_frac_signed_per_threshold": {
507
+ "0.5": 0.9982638888888888,
508
+ "0.6": 1.0,
509
+ "0.7": 1.0,
510
+ "0.8": 1.0,
511
+ "0.9": 1.0
512
+ },
513
+ "unique_frac_signed_band": 1.0,
514
+ "unique_frac_signed_band_std": 0.0,
515
+ "matched_frac_signed_per_threshold": {
516
+ "0.5": 0.9004991319444443,
517
+ "0.6": 0.8782552083333334,
518
+ "0.7": 0.8550347222222222,
519
+ "0.8": 0.8266059027777777,
520
+ "0.9": 0.7838541666666666
521
+ },
522
+ "matched_frac_signed_band": 0.8359374999999999,
523
+ "matched_frac_signed_band_std": 0.0034641596116198664,
524
+ "babel_mu1_abs_mean": 0.20671896636486053,
525
+ "babel_mu1_abs_std": 0.0005238186023812209,
526
+ "n_dup_abs_per_threshold": {
527
+ "0.5": 49.333333333333336,
528
+ "0.6": 24.0,
529
+ "0.7": 12.0,
530
+ "0.8": 3.3333333333333335,
531
+ "0.9": 1.3333333333333333
532
+ },
533
+ "n_dup_abs_band": 10.166666666666668,
534
+ "n_dup_abs_band_std": 2.3094010767585034,
535
+ "n_dup_clusters_abs_per_threshold": {
536
+ "0.5": 23.0,
537
+ "0.6": 12.0,
538
+ "0.7": 6.0,
539
+ "0.8": 1.6666666666666667,
540
+ "0.9": 0.6666666666666666
541
+ },
542
+ "n_dup_clusters_abs_band": 5.083333333333334,
543
+ "n_dup_clusters_abs_band_std": 1.1547005383792517,
544
+ "largest_dup_abs_per_threshold": {
545
+ "0.5": 4.333333333333333,
546
+ "0.6": 2.0,
547
+ "0.7": 2.0,
548
+ "0.8": 2.0,
549
+ "0.9": 1.6666666666666667
550
+ },
551
+ "largest_dup_abs_band": 1.9166666666666667,
552
+ "largest_dup_abs_band_std": 0.14433756729740646,
553
+ "dcs_abs_mean": 0.8827048540115356,
554
+ "dcs_abs_std": 0.001661941476915841,
555
+ "dedup_dcs_abs_per_threshold": {
556
+ "0.5": 0.8771897977696628,
557
+ "0.6": 0.8807891646219005,
558
+ "0.7": 0.8818634386924612,
559
+ "0.8": 0.882397524103807,
560
+ "0.9": 0.8827307641040534
561
+ },
562
+ "dedup_dcs_abs_band": 0.8819452228805555,
563
+ "dedup_dcs_abs_band_std": 0.0017468567720698796,
564
+ "matched_cos_abs_per_threshold": {
565
+ "0.5": 0.8847741087277731,
566
+ "0.6": 0.8842431704203287,
567
+ "0.7": 0.8835894266764323,
568
+ "0.8": 0.8828764160474142,
569
+ "0.9": 0.882922371228536
570
+ },
571
+ "matched_cos_abs_band": 0.8834078460931778,
572
+ "matched_cos_abs_band_std": 0.0017490432337512521,
573
+ "unique_frac_abs_per_threshold": {
574
+ "0.5": 0.9914279513888888,
575
+ "0.6": 0.99609375,
576
+ "0.7": 0.998046875,
577
+ "0.8": 0.9994574652777778,
578
+ "0.9": 0.9997829861111112
579
+ },
580
+ "unique_frac_abs_band": 0.9983452690972222,
581
+ "unique_frac_abs_band_std": 0.00018793954075191677,
582
+ "matched_frac_abs_per_threshold": {
583
+ "0.5": 0.896484375,
584
+ "0.6": 0.8767361111111112,
585
+ "0.7": 0.8544921875,
586
+ "0.8": 0.826388888888889,
587
+ "0.9": 0.7838541666666666
588
+ },
589
+ "matched_frac_abs_band": 0.8353678385416666,
590
+ "matched_frac_abs_band_std": 0.0033484766620183524,
591
+ "alive_source": "stats",
592
+ "frac_denominator": "d_sae",
593
+ "d_sae": 3072
594
+ }
595
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 0.74,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.0668,
3
+ "r2_score": 0.9819,
4
+ "l0_per_token_0": 0.0448,
5
+ "l0_per_token_0.001": 0.0448,
6
+ "l0_per_token_0.01": 0.0448,
7
+ "l0_per_token_0.05": 0.0447,
8
+ "l0_per_token_0.1": 0.0447,
9
+ "l0_per_token_0.5": 0.0303,
10
+ "l0_per_token_1.0": 0.0118,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0016,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 150.1596
15
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 0.74,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.0667,
3
+ "r2_score": 0.9819,
4
+ "l0_per_token_0": 0.0449,
5
+ "l0_per_token_0.001": 0.0449,
6
+ "l0_per_token_0.01": 0.0449,
7
+ "l0_per_token_0.05": 0.0449,
8
+ "l0_per_token_0.1": 0.0449,
9
+ "l0_per_token_0.5": 0.0302,
10
+ "l0_per_token_1.0": 0.0118,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0013,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 149.8765
15
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 0.74,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.0667,
3
+ "r2_score": 0.9819,
4
+ "l0_per_token_0": 0.0451,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0016,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 151.3502
9
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 31.6107626459144,
20
+ "recon_r2": 0.9484165452418616,
21
+ "l1_per_token": 56.83691613633403
22
+ },
23
+ "seed1": {
24
+ "l0": 31.69722568093385,
25
+ "recon_r2": 0.9484495701389029,
26
+ "l1_per_token": 56.897462397093484
27
+ },
28
+ "seed2": {
29
+ "l0": 31.602,
30
+ "recon_r2": 0.9484986321449965,
31
+ "l1_per_token": 57.05980449782456
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.868192830147582,
38
+ "mcc_abs": 0.8682154949809581,
39
+ "act_iou": 0.6602800140067636,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.995483385236984,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.9222666383001389,
44
+ "0.7": 0.9347953888839434,
45
+ "0.8": 0.9482348602195857,
46
+ "0.9": 0.9666846755577978
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.9217468605917245,
50
+ "0.7": 0.9348605627095109,
51
+ "0.8": 0.9481281505391772,
52
+ "0.9": 0.9666214681910065
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.8627724811817782,
58
+ "mcc_abs": 0.862792552246459,
59
+ "act_iou": 0.6552409996875739,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9953709878345616,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.9215142897913385,
64
+ "0.7": 0.9332173589593314,
65
+ "0.8": 0.9489304320031606,
66
+ "0.9": 0.9682569110224258
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.922274592562494,
70
+ "0.7": 0.9333504700494636,
71
+ "0.8": 0.9489673250368282,
72
+ "0.9": 0.9682246831921323
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.8681324772026745,
78
+ "mcc_abs": 0.8681471772515105,
79
+ "act_iou": 0.6598968191338653,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.995452721178881,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.924551068018861,
84
+ "0.7": 0.9357995175703497,
85
+ "0.8": 0.95156812792084,
86
+ "0.9": 0.9707907761224103
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.924748055992792,
90
+ "0.7": 0.9357541937666937,
91
+ "0.8": 0.9517322793592582,
92
+ "0.9": 0.9707856830970484
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 31.636662775616085,
106
+ "l0_std": 0.052631693077954025,
107
+ "recon_r2_mean": 0.9484549158419203,
108
+ "recon_r2_std": 4.1303720439378714e-05,
109
+ "l1_per_token_mean": 56.931394343750696,
110
+ "l1_per_token_std": 0.11525336513500806,
111
+ "input_r2_mean": 0.9954356980834755,
112
+ "input_r2_std": 5.8100200952439126e-05,
113
+ "mcc_signed_mean": 0.8663659295106783,
114
+ "mcc_signed_std": 0.003112163843509142,
115
+ "mcc_abs_mean": 0.8663850748263092,
116
+ "mcc_abs_std": 0.0031114033317924607,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.9227773320367795,
119
+ "0.7": 0.9346040884712082,
120
+ "0.8": 0.9495778067145287,
121
+ "0.9": 0.9685774542342114
122
+ },
123
+ "matched_mcc_signed_band": 0.9438841703641819,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.9229231697156702,
126
+ "0.7": 0.9346550755085561,
127
+ "0.8": 0.9496092516450879,
128
+ "0.9": 0.9685439448267291
129
+ },
130
+ "matched_mcc_abs_band": 0.9439328604240109,
131
+ "act_iou_mean": 0.6584726109427343,
132
+ "act_iou_std": 0.0028052082019011356,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 3072,
20
+ "d_in": 768,
21
+ "l0": 32,
22
+ "l0_raw": 31.6416,
23
+ "recon_r2": 0.9507,
24
+ "l1_per_token": 56.8797,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 3,
30
+ "dense_rate": 0.0009765625,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 255,
34
+ "0.6": 0,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 124,
41
+ "0.6": 0,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 3,
48
+ "0.6": 1,
49
+ "0.7": 1,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.3518010079860687,
54
+ "n_dup_abs": {
55
+ "0.5": 268,
56
+ "0.6": 12,
57
+ "0.7": 4,
58
+ "0.8": 0,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 131,
63
+ "0.6": 6,
64
+ "0.7": 2,
65
+ "0.8": 0,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 3,
70
+ "0.6": 2,
71
+ "0.7": 2,
72
+ "0.8": 1,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.3539195954799652
76
+ },
77
+ "seed1": {
78
+ "d_sae": 3072,
79
+ "d_in": 768,
80
+ "l0": 32,
81
+ "l0_raw": 31.6416,
82
+ "recon_r2": 0.9507,
83
+ "l1_per_token": 56.9402,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 1,
89
+ "dense_rate": 0.0003255208333333333,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 275,
93
+ "0.6": 0,
94
+ "0.7": 0,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 134,
100
+ "0.6": 0,
101
+ "0.7": 0,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 4,
107
+ "0.6": 1,
108
+ "0.7": 1,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.35239484906196594,
113
+ "n_dup_abs": {
114
+ "0.5": 294,
115
+ "0.6": 8,
116
+ "0.7": 2,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 144,
122
+ "0.6": 4,
123
+ "0.7": 1,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 3,
129
+ "0.6": 2,
130
+ "0.7": 2,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.35446372628211975
135
+ },
136
+ "seed2": {
137
+ "d_sae": 3072,
138
+ "d_in": 768,
139
+ "l0": 32,
140
+ "l0_raw": 31.6416,
141
+ "recon_r2": 0.9485,
142
+ "l1_per_token": 57.0986,
143
+ "stats_source": "stats",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 2,
148
+ "dense_rate": 0.0006510416666666666,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 283,
152
+ "0.6": 0,
153
+ "0.7": 0,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 138,
159
+ "0.6": 0,
160
+ "0.7": 0,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 4,
166
+ "0.6": 1,
167
+ "0.7": 1,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.35185420513153076,
172
+ "n_dup_abs": {
173
+ "0.5": 291,
174
+ "0.6": 12,
175
+ "0.7": 4,
176
+ "0.8": 0,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 141,
181
+ "0.6": 6,
182
+ "0.7": 2,
183
+ "0.8": 0,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 4,
188
+ "0.6": 2,
189
+ "0.7": 2,
190
+ "0.8": 1,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.3537638187408447
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.9063729643821716,
199
+ "dcs_abs": 0.9064735770225525,
200
+ "frac_sign_flipped": 0.0048828125,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 3072,
204
+ "n_alive_b": 3072,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.9115521907806396,
207
+ "0.6": 0.9063729643821716,
208
+ "0.7": 0.9063729643821716,
209
+ "0.8": 0.9063729643821716,
210
+ "0.9": 0.9063729643821716
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.9557291666666666,
214
+ "0.6": 1.0,
215
+ "0.7": 1.0,
216
+ "0.8": 1.0,
217
+ "0.9": 1.0
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.871197015667955,
221
+ "0.6": 0.9063729643821716,
222
+ "0.7": 0.9063729643821716,
223
+ "0.8": 0.9063729643821716,
224
+ "0.9": 0.9063729643821716
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.9013671875,
228
+ "0.6": 0.9270833333333334,
229
+ "0.7": 0.9007161458333334,
230
+ "0.8": 0.8642578125,
231
+ "0.9": 0.7835286458333334
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.9123555421829224,
235
+ "0.6": 0.9068666696548462,
236
+ "0.7": 0.9066073894500732,
237
+ "0.8": 0.9064735770225525,
238
+ "0.9": 0.9064735770225525
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.9532877604166666,
242
+ "0.6": 0.9983723958333334,
243
+ "0.7": 0.99951171875,
244
+ "0.8": 1.0,
245
+ "0.9": 1.0
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.8697373715112917,
249
+ "0.6": 0.9053906496847048,
250
+ "0.7": 0.9061647100606933,
251
+ "0.8": 0.9064735770225525,
252
+ "0.9": 0.9064735770225525
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.8987630208333334,
256
+ "0.6": 0.9251302083333334,
257
+ "0.7": 0.8990885416666666,
258
+ "0.8": 0.8645833333333334,
259
+ "0.9": 0.7838541666666666
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.9020346999168396,
264
+ "dcs_abs": 0.9021320343017578,
265
+ "frac_sign_flipped": 0.004557291666666667,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 3072,
269
+ "n_alive_b": 3072,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.9078397154808044,
272
+ "0.6": 0.9020347595214844,
273
+ "0.7": 0.9020347595214844,
274
+ "0.8": 0.9020347595214844,
275
+ "0.9": 0.9020347595214844
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.955078125,
279
+ "0.6": 1.0,
280
+ "0.7": 1.0,
281
+ "0.8": 1.0,
282
+ "0.9": 1.0
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.8670578532619402,
286
+ "0.6": 0.9020347595214844,
287
+ "0.7": 0.9020347595214844,
288
+ "0.8": 0.9020347595214844,
289
+ "0.9": 0.9020347595214844
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.8974609375,
293
+ "0.6": 0.9192708333333334,
294
+ "0.7": 0.896484375,
295
+ "0.8": 0.8570963541666666,
296
+ "0.9": 0.7779947916666666
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.9080588221549988,
300
+ "0.6": 0.9023743271827698,
301
+ "0.7": 0.9021952748298645,
302
+ "0.8": 0.902131974697113,
303
+ "0.9": 0.902131974697113
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.9532877604166666,
307
+ "0.6": 0.998046875,
308
+ "0.7": 0.9993489583333334,
309
+ "0.8": 1.0,
310
+ "0.9": 1.0
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.865641360898735,
314
+ "0.6": 0.9006118773249909,
315
+ "0.7": 0.9016079081144805,
316
+ "0.8": 0.902131974697113,
317
+ "0.9": 0.902131974697113
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.8955078125,
321
+ "0.6": 0.9173177083333334,
322
+ "0.7": 0.8961588541666666,
323
+ "0.8": 0.8567708333333334,
324
+ "0.9": 0.7779947916666666
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.9052200317382812,
329
+ "dcs_abs": 0.9053223729133606,
330
+ "frac_sign_flipped": 0.0048828125,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 3072,
334
+ "n_alive_b": 3072,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.908517062664032,
337
+ "0.6": 0.9052199721336365,
338
+ "0.7": 0.9052199721336365,
339
+ "0.8": 0.9052199721336365,
340
+ "0.9": 0.9052199721336365
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.9534505208333334,
344
+ "0.6": 1.0,
345
+ "0.7": 1.0,
346
+ "0.8": 1.0,
347
+ "0.9": 1.0
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.8662260665829915,
351
+ "0.6": 0.9052199721336365,
352
+ "0.7": 0.9052199721336365,
353
+ "0.8": 0.9052199721336365,
354
+ "0.9": 0.9052199721336365
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.8961588541666666,
358
+ "0.6": 0.9235026041666666,
359
+ "0.7": 0.900390625,
360
+ "0.8": 0.8600260416666666,
361
+ "0.9": 0.78125
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.9079200625419617,
365
+ "0.6": 0.9056766629219055,
366
+ "0.7": 0.9055261611938477,
367
+ "0.8": 0.9053223729133606,
368
+ "0.9": 0.9053223729133606
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.951171875,
372
+ "0.6": 0.9983723958333334,
373
+ "0.7": 0.99951171875,
374
+ "0.8": 1.0,
375
+ "0.9": 1.0
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.863588028238155,
379
+ "0.6": 0.9042025798116811,
380
+ "0.7": 0.9050840097479522,
381
+ "0.8": 0.9053223729133606,
382
+ "0.9": 0.9053223729133606
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.8929036458333334,
386
+ "0.6": 0.9228515625,
387
+ "0.7": 0.9000651041666666,
388
+ "0.8": 0.8597005208333334,
389
+ "0.9": 0.7815755208333334
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 0,
403
+ "n_dead_std": 0.0,
404
+ "n_dense_mean": 2,
405
+ "n_dense_std": 1.0,
406
+ "l0_mean": 32,
407
+ "l0_std": 0.0,
408
+ "recon_r2_mean": 0.9499666666666666,
409
+ "recon_r2_std": 0.001270170592217165,
410
+ "l1_per_token_mean": 56.972833333333334,
411
+ "l1_per_token_std": 0.11303983073825415,
412
+ "babel_mu1_signed_mean": 0.3520166873931885,
413
+ "babel_mu1_signed_std": 0.00032857597275325165,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 271.0,
416
+ "0.6": 0.0,
417
+ "0.7": 0.0,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 0.0,
422
+ "n_dup_signed_band_std": 0.0,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 132.0,
425
+ "0.6": 0.0,
426
+ "0.7": 0.0,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 0.0,
431
+ "n_dup_clusters_signed_band_std": 0.0,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 3.6666666666666665,
434
+ "0.6": 1.0,
435
+ "0.7": 1.0,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 1.0,
440
+ "largest_dup_signed_band_std": 0.0,
441
+ "dcs_signed_mean": 0.9045425653457642,
442
+ "dcs_signed_std": 0.0022470770297284524,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.8681603118376288,
445
+ "0.6": 0.9045425653457642,
446
+ "0.7": 0.9045425653457642,
447
+ "0.8": 0.9045425653457642,
448
+ "0.9": 0.9045425653457642
449
+ },
450
+ "dedup_dcs_signed_band": 0.9045425653457642,
451
+ "dedup_dcs_signed_band_std": 0.002247034784001852,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.9093029896418253,
454
+ "0.6": 0.9045425653457642,
455
+ "0.7": 0.9045425653457642,
456
+ "0.8": 0.9045425653457642,
457
+ "0.9": 0.9045425653457642
458
+ },
459
+ "matched_cos_signed_band": 0.9045425653457642,
460
+ "matched_cos_signed_band_std": 0.002247034784001852,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.9547526041666666,
463
+ "0.6": 1.0,
464
+ "0.7": 1.0,
465
+ "0.8": 1.0,
466
+ "0.9": 1.0
467
+ },
468
+ "unique_frac_signed_band": 1.0,
469
+ "unique_frac_signed_band_std": 0.0,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.8983289930555555,
472
+ "0.6": 0.9232855902777778,
473
+ "0.7": 0.8991970486111112,
474
+ "0.8": 0.8604600694444443,
475
+ "0.9": 0.7809244791666666
476
+ },
477
+ "matched_frac_signed_band": 0.8659667968749999,
478
+ "matched_frac_signed_band_std": 0.0031052708379458544,
479
+ "babel_mu1_abs_mean": 0.3540490468343099,
480
+ "babel_mu1_abs_std": 0.00036747228662807694,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 284.3333333333333,
483
+ "0.6": 10.666666666666666,
484
+ "0.7": 3.3333333333333335,
485
+ "0.8": 0.0,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 3.5,
489
+ "n_dup_abs_band_std": 0.8660254037844386,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 138.66666666666666,
492
+ "0.6": 5.333333333333333,
493
+ "0.7": 1.6666666666666667,
494
+ "0.8": 0.0,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 1.75,
498
+ "n_dup_clusters_abs_band_std": 0.4330127018922193,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 3.3333333333333335,
501
+ "0.6": 2.0,
502
+ "0.7": 2.0,
503
+ "0.8": 1.0,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 1.5,
507
+ "largest_dup_abs_band_std": 0.0,
508
+ "dcs_abs_mean": 0.904642661412557,
509
+ "dcs_abs_std": 0.0022491674220397324,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.8663222535493939,
512
+ "0.6": 0.9034017022737922,
513
+ "0.7": 0.9042855426410421,
514
+ "0.8": 0.904642641544342,
515
+ "0.9": 0.904642641544342
516
+ },
517
+ "dedup_dcs_abs_band": 0.9042431320008796,
518
+ "dedup_dcs_abs_band_std": 0.002341674142308456,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.9094448089599609,
521
+ "0.6": 0.9049725532531738,
522
+ "0.7": 0.9047762751579285,
523
+ "0.8": 0.904642641544342,
524
+ "0.9": 0.904642641544342
525
+ },
526
+ "matched_cos_abs_band": 0.9047585278749466,
527
+ "matched_cos_abs_band_std": 0.002281284214874081,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.9525824652777777,
530
+ "0.6": 0.998263888888889,
531
+ "0.7": 0.9994574652777778,
532
+ "0.8": 1.0,
533
+ "0.9": 1.0
534
+ },
535
+ "unique_frac_abs_band": 0.9994303385416667,
536
+ "unique_frac_abs_band_std": 7.047732778193675e-05,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.895724826388889,
539
+ "0.6": 0.9217664930555557,
540
+ "0.7": 0.8984375,
541
+ "0.8": 0.8603515625,
542
+ "0.9": 0.7811414930555555
543
+ },
544
+ "matched_frac_abs_band": 0.8654242621527778,
545
+ "matched_frac_abs_band_std": 0.003099222180899978,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 3072
549
+ }
550
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.76,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1905,
3
+ "r2_score": 0.9507,
4
+ "l0_per_token_0": 0.0103,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 56.8797
9
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.76,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1903,
3
+ "r2_score": 0.9507,
4
+ "l0_per_token_0": 0.0103,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 56.9402
9
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.76,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1901,
3
+ "r2_score": 0.9485,
4
+ "l0_per_token_0": 0.0103,
5
+ "l0_per_token_0.001": 0.0103,
6
+ "l0_per_token_0.01": 0.0103,
7
+ "l0_per_token_0.05": 0.0103,
8
+ "l0_per_token_0.1": 0.0103,
9
+ "l0_per_token_0.5": 0.0076,
10
+ "l0_per_token_1.0": 0.0049,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 57.0986
15
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 68.03946692607003,
20
+ "recon_r2": 0.9696529642250379,
21
+ "l1_per_token": 104.7458964975562
22
+ },
23
+ "seed1": {
24
+ "l0": 67.2137328145266,
25
+ "recon_r2": 0.9687827232591277,
26
+ "l1_per_token": 101.17873846409823
27
+ },
28
+ "seed2": {
29
+ "l0": 64.8874513618677,
30
+ "recon_r2": 0.9680454074676013,
31
+ "l1_per_token": 96.49608938424814
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.8571498961907832,
38
+ "mcc_abs": 0.857158631942457,
39
+ "act_iou": 0.5562796591080029,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.9927267101828172,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.9249115673201719,
44
+ "0.7": 0.9337020850750817,
45
+ "0.8": 0.9462472454367687,
46
+ "0.9": 0.9575276936903444
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.9245309046523821,
50
+ "0.7": 0.9336897316895995,
51
+ "0.8": 0.9459720278700849,
52
+ "0.9": 0.9577268962523132
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.8673294933679094,
58
+ "mcc_abs": 0.867345856620784,
59
+ "act_iou": 0.5669167639805206,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9929314655368502,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.9307832303934402,
64
+ "0.7": 0.9380695497828306,
65
+ "0.8": 0.9469018408421415,
66
+ "0.9": 0.9587488587728923
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.9306730982487056,
70
+ "0.7": 0.9380485685272566,
71
+ "0.8": 0.9469496767339759,
72
+ "0.9": 0.9587488587728926
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.8611214823312183,
78
+ "mcc_abs": 0.8611404155722532,
79
+ "act_iou": 0.5623300066505211,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.9932440410528658,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.9278395149961061,
84
+ "0.7": 0.9375869317884058,
85
+ "0.8": 0.9497104273968842,
86
+ "0.9": 0.9617156732997467
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.9281774340929967,
90
+ "0.7": 0.9374782302280529,
91
+ "0.8": 0.9490106833283541,
92
+ "0.9": 0.9617296637209065
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 66.71355036748811,
106
+ "l0_std": 1.6344532386971145,
107
+ "recon_r2_mean": 0.9688270316505889,
108
+ "recon_r2_std": 0.0008046937972051536,
109
+ "l1_per_token_mean": 100.80690811530086,
110
+ "l1_per_token_std": 4.137453650243368,
111
+ "input_r2_mean": 0.9929674055908443,
112
+ "input_r2_std": 0.0002605313280339487,
113
+ "mcc_signed_mean": 0.861866957296637,
114
+ "mcc_signed_std": 0.005130579827386824,
115
+ "mcc_abs_mean": 0.8618816347118314,
116
+ "mcc_abs_std": 0.005133901150327738,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.9278447709032394,
119
+ "0.7": 0.9364528555487727,
120
+ "0.8": 0.9476198378919314,
121
+ "0.9": 0.9593307419209944
122
+ },
123
+ "matched_mcc_signed_band": 0.9428120515662345,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.9277938123313615,
126
+ "0.7": 0.9364055101483029,
127
+ "0.8": 0.9473107959774717,
128
+ "0.9": 0.9594018062487041
129
+ },
130
+ "matched_mcc_abs_band": 0.9427279811764601,
131
+ "act_iou_mean": 0.5618421432463482,
132
+ "act_iou_std": 0.005335307680263924,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/dict_metrics_step_final.json ADDED
@@ -0,0 +1,595 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 3,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 3072,
20
+ "d_in": 768,
21
+ "l0": 67,
22
+ "l0_raw": 67.2768,
23
+ "recon_r2": 0.971,
24
+ "l1_per_token": 103.8579,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 14,
30
+ "dense_rate": 0.004557291666666667,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 51,
34
+ "0.6": 0,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 25,
41
+ "0.6": 0,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "n_dup_removed_signed": {
47
+ "0.5": 26,
48
+ "0.6": 0,
49
+ "0.7": 0,
50
+ "0.8": 0,
51
+ "0.9": 0
52
+ },
53
+ "largest_dup_signed": {
54
+ "0.5": 3,
55
+ "0.6": 1,
56
+ "0.7": 1,
57
+ "0.8": 1,
58
+ "0.9": 1
59
+ },
60
+ "babel_mu1_signed": 0.24493642151355743,
61
+ "n_dup_abs": {
62
+ "0.5": 74,
63
+ "0.6": 10,
64
+ "0.7": 6,
65
+ "0.8": 2,
66
+ "0.9": 0
67
+ },
68
+ "n_dup_clusters_abs": {
69
+ "0.5": 36,
70
+ "0.6": 5,
71
+ "0.7": 3,
72
+ "0.8": 1,
73
+ "0.9": 0
74
+ },
75
+ "n_dup_removed_abs": {
76
+ "0.5": 38,
77
+ "0.6": 5,
78
+ "0.7": 3,
79
+ "0.8": 1,
80
+ "0.9": 0
81
+ },
82
+ "largest_dup_abs": {
83
+ "0.5": 3,
84
+ "0.6": 2,
85
+ "0.7": 2,
86
+ "0.8": 2,
87
+ "0.9": 1
88
+ },
89
+ "babel_mu1_abs": 0.2513481378555298,
90
+ "dup_excludes_dead": true
91
+ },
92
+ "seed1": {
93
+ "d_sae": 3072,
94
+ "d_in": 768,
95
+ "l0": 67,
96
+ "l0_raw": 67.2768,
97
+ "recon_r2": 0.9702,
98
+ "l1_per_token": 101.2308,
99
+ "stats_source": "stats",
100
+ "n_dead": 0,
101
+ "dead_rate": 0.0,
102
+ "alive_source": "stats",
103
+ "n_dense": 9,
104
+ "dense_rate": 0.0029296875,
105
+ "dense_threshold": 0.1,
106
+ "n_dup_signed": {
107
+ "0.5": 64,
108
+ "0.6": 0,
109
+ "0.7": 0,
110
+ "0.8": 0,
111
+ "0.9": 0
112
+ },
113
+ "n_dup_clusters_signed": {
114
+ "0.5": 32,
115
+ "0.6": 0,
116
+ "0.7": 0,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_removed_signed": {
121
+ "0.5": 32,
122
+ "0.6": 0,
123
+ "0.7": 0,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_signed": {
128
+ "0.5": 2,
129
+ "0.6": 1,
130
+ "0.7": 1,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_signed": 0.2563404142856598,
135
+ "n_dup_abs": {
136
+ "0.5": 73,
137
+ "0.6": 8,
138
+ "0.7": 4,
139
+ "0.8": 2,
140
+ "0.9": 0
141
+ },
142
+ "n_dup_clusters_abs": {
143
+ "0.5": 35,
144
+ "0.6": 4,
145
+ "0.7": 2,
146
+ "0.8": 1,
147
+ "0.9": 0
148
+ },
149
+ "n_dup_removed_abs": {
150
+ "0.5": 38,
151
+ "0.6": 4,
152
+ "0.7": 2,
153
+ "0.8": 1,
154
+ "0.9": 0
155
+ },
156
+ "largest_dup_abs": {
157
+ "0.5": 4,
158
+ "0.6": 2,
159
+ "0.7": 2,
160
+ "0.8": 2,
161
+ "0.9": 1
162
+ },
163
+ "babel_mu1_abs": 0.26287007331848145,
164
+ "dup_excludes_dead": true
165
+ },
166
+ "seed2": {
167
+ "d_sae": 3072,
168
+ "d_in": 768,
169
+ "l0": 65,
170
+ "l0_raw": 64.8192,
171
+ "recon_r2": 0.968,
172
+ "l1_per_token": 96.535,
173
+ "stats_source": "stats",
174
+ "n_dead": 0,
175
+ "dead_rate": 0.0,
176
+ "alive_source": "stats",
177
+ "n_dense": 12,
178
+ "dense_rate": 0.00390625,
179
+ "dense_threshold": 0.1,
180
+ "n_dup_signed": {
181
+ "0.5": 70,
182
+ "0.6": 0,
183
+ "0.7": 0,
184
+ "0.8": 0,
185
+ "0.9": 0
186
+ },
187
+ "n_dup_clusters_signed": {
188
+ "0.5": 34,
189
+ "0.6": 0,
190
+ "0.7": 0,
191
+ "0.8": 0,
192
+ "0.9": 0
193
+ },
194
+ "n_dup_removed_signed": {
195
+ "0.5": 36,
196
+ "0.6": 0,
197
+ "0.7": 0,
198
+ "0.8": 0,
199
+ "0.9": 0
200
+ },
201
+ "largest_dup_signed": {
202
+ "0.5": 3,
203
+ "0.6": 1,
204
+ "0.7": 1,
205
+ "0.8": 1,
206
+ "0.9": 1
207
+ },
208
+ "babel_mu1_signed": 0.2591460645198822,
209
+ "n_dup_abs": {
210
+ "0.5": 91,
211
+ "0.6": 12,
212
+ "0.7": 6,
213
+ "0.8": 2,
214
+ "0.9": 0
215
+ },
216
+ "n_dup_clusters_abs": {
217
+ "0.5": 43,
218
+ "0.6": 6,
219
+ "0.7": 3,
220
+ "0.8": 1,
221
+ "0.9": 0
222
+ },
223
+ "n_dup_removed_abs": {
224
+ "0.5": 48,
225
+ "0.6": 6,
226
+ "0.7": 3,
227
+ "0.8": 1,
228
+ "0.9": 0
229
+ },
230
+ "largest_dup_abs": {
231
+ "0.5": 4,
232
+ "0.6": 2,
233
+ "0.7": 2,
234
+ "0.8": 2,
235
+ "0.9": 1
236
+ },
237
+ "babel_mu1_abs": 0.2650682032108307,
238
+ "dup_excludes_dead": true
239
+ }
240
+ },
241
+ "per_pair": {
242
+ "seed0\u2194seed1": {
243
+ "dcs_signed": 0.897960901260376,
244
+ "dcs_abs": 0.8985502123832703,
245
+ "frac_sign_flipped": 0.010416666666666666,
246
+ "alive_source": "stats",
247
+ "frac_denominator": "d_sae",
248
+ "n_alive_a": 3072,
249
+ "n_alive_b": 3072,
250
+ "matched_cos_signed": {
251
+ "0.5": 0.8992840647697449,
252
+ "0.6": 0.8979608416557312,
253
+ "0.7": 0.8979608416557312,
254
+ "0.8": 0.8979608416557312,
255
+ "0.9": 0.8979608416557312
256
+ },
257
+ "unique_frac_signed": {
258
+ "0.5": 0.9905598958333334,
259
+ "0.6": 1.0,
260
+ "0.7": 1.0,
261
+ "0.8": 1.0,
262
+ "0.9": 1.0
263
+ },
264
+ "dedup_dcs_signed": {
265
+ "0.5": 0.8907947295228951,
266
+ "0.6": 0.8979608416557312,
267
+ "0.7": 0.8979608416557312,
268
+ "0.8": 0.8979608416557312,
269
+ "0.9": 0.8979608416557312
270
+ },
271
+ "matched_frac_signed": {
272
+ "0.5": 0.9098307291666666,
273
+ "0.6": 0.9052734375,
274
+ "0.7": 0.8873697916666666,
275
+ "0.8": 0.853515625,
276
+ "0.9": 0.8059895833333334
277
+ },
278
+ "matched_cos_abs": {
279
+ "0.5": 0.8993204236030579,
280
+ "0.6": 0.898636519908905,
281
+ "0.7": 0.8987310528755188,
282
+ "0.8": 0.8985503315925598,
283
+ "0.9": 0.8985502123832703
284
+ },
285
+ "unique_frac_abs": {
286
+ "0.5": 0.9876302083333334,
287
+ "0.6": 0.99853515625,
288
+ "0.7": 0.9991861979166666,
289
+ "0.8": 0.9996744791666666,
290
+ "0.9": 1.0
291
+ },
292
+ "dedup_dcs_abs": {
293
+ "0.5": 0.8881960173215097,
294
+ "0.6": 0.8973201578191947,
295
+ "0.7": 0.8979996636723323,
296
+ "0.8": 0.8982578347398278,
297
+ "0.9": 0.8985502123832703
298
+ },
299
+ "matched_frac_abs": {
300
+ "0.5": 0.9098307291666666,
301
+ "0.6": 0.9046223958333334,
302
+ "0.7": 0.8873697916666666,
303
+ "0.8": 0.853515625,
304
+ "0.9": 0.8056640625
305
+ }
306
+ },
307
+ "seed0\u2194seed2": {
308
+ "dcs_signed": 0.9055841565132141,
309
+ "dcs_abs": 0.905812680721283,
310
+ "frac_sign_flipped": 0.008138020833333334,
311
+ "alive_source": "stats",
312
+ "frac_denominator": "d_sae",
313
+ "n_alive_a": 3072,
314
+ "n_alive_b": 3072,
315
+ "matched_cos_signed": {
316
+ "0.5": 0.9075362682342529,
317
+ "0.6": 0.9055840969085693,
318
+ "0.7": 0.9055840969085693,
319
+ "0.8": 0.9055840969085693,
320
+ "0.9": 0.9055840969085693
321
+ },
322
+ "unique_frac_signed": {
323
+ "0.5": 0.9899088541666666,
324
+ "0.6": 1.0,
325
+ "0.7": 1.0,
326
+ "0.8": 1.0,
327
+ "0.9": 1.0
328
+ },
329
+ "dedup_dcs_signed": {
330
+ "0.5": 0.8983781874024619,
331
+ "0.6": 0.9055840969085693,
332
+ "0.7": 0.9055840969085693,
333
+ "0.8": 0.9055840969085693,
334
+ "0.9": 0.9055840969085693
335
+ },
336
+ "matched_frac_signed": {
337
+ "0.5": 0.9186197916666666,
338
+ "0.6": 0.9108072916666666,
339
+ "0.7": 0.8951822916666666,
340
+ "0.8": 0.869140625,
341
+ "0.9": 0.818359375
342
+ },
343
+ "matched_cos_abs": {
344
+ "0.5": 0.9082745909690857,
345
+ "0.6": 0.905983567237854,
346
+ "0.7": 0.9058568477630615,
347
+ "0.8": 0.9058550000190735,
348
+ "0.9": 0.9058125615119934
349
+ },
350
+ "unique_frac_abs": {
351
+ "0.5": 0.9860026041666666,
352
+ "0.6": 0.9982096354166666,
353
+ "0.7": 0.9990234375,
354
+ "0.8": 0.9996744791666666,
355
+ "0.9": 1.0
356
+ },
357
+ "dedup_dcs_abs": {
358
+ "0.5": 0.8955611119939324,
359
+ "0.6": 0.9043615263459893,
360
+ "0.7": 0.9049722219351679,
361
+ "0.8": 0.9055601253445881,
362
+ "0.9": 0.9058125615119934
363
+ },
364
+ "matched_frac_abs": {
365
+ "0.5": 0.916015625,
366
+ "0.6": 0.9085286458333334,
367
+ "0.7": 0.89453125,
368
+ "0.8": 0.8694661458333334,
369
+ "0.9": 0.818359375
370
+ }
371
+ },
372
+ "seed1\u2194seed2": {
373
+ "dcs_signed": 0.8994863629341125,
374
+ "dcs_abs": 0.9001209735870361,
375
+ "frac_sign_flipped": 0.0107421875,
376
+ "alive_source": "stats",
377
+ "frac_denominator": "d_sae",
378
+ "n_alive_a": 3072,
379
+ "n_alive_b": 3072,
380
+ "matched_cos_signed": {
381
+ "0.5": 0.9010207653045654,
382
+ "0.6": 0.8994863629341125,
383
+ "0.7": 0.8994863629341125,
384
+ "0.8": 0.8994863629341125,
385
+ "0.9": 0.8994863629341125
386
+ },
387
+ "unique_frac_signed": {
388
+ "0.5": 0.9889322916666666,
389
+ "0.6": 1.0,
390
+ "0.7": 1.0,
391
+ "0.8": 1.0,
392
+ "0.9": 1.0
393
+ },
394
+ "dedup_dcs_signed": {
395
+ "0.5": 0.8910485302718977,
396
+ "0.6": 0.8994863629341125,
397
+ "0.7": 0.8994863629341125,
398
+ "0.8": 0.8994863629341125,
399
+ "0.9": 0.8994863629341125
400
+ },
401
+ "matched_frac_signed": {
402
+ "0.5": 0.9098307291666666,
403
+ "0.6": 0.9049479166666666,
404
+ "0.7": 0.8834635416666666,
405
+ "0.8": 0.8512369791666666,
406
+ "0.9": 0.8040364583333334
407
+ },
408
+ "matched_cos_abs": {
409
+ "0.5": 0.9026309847831726,
410
+ "0.6": 0.9004991054534912,
411
+ "0.7": 0.9003547430038452,
412
+ "0.8": 0.9001197218894958,
413
+ "0.9": 0.9001209735870361
414
+ },
415
+ "unique_frac_abs": {
416
+ "0.5": 0.9860026041666666,
417
+ "0.6": 0.9983723958333334,
418
+ "0.7": 0.9991861979166666,
419
+ "0.8": 0.9996744791666666,
420
+ "0.9": 1.0
421
+ },
422
+ "dedup_dcs_abs": {
423
+ "0.5": 0.889996501597731,
424
+ "0.6": 0.8990334493573755,
425
+ "0.7": 0.8996220324382496,
426
+ "0.8": 0.8998267141675266,
427
+ "0.9": 0.9001209735870361
428
+ },
429
+ "matched_frac_abs": {
430
+ "0.5": 0.9078776041666666,
431
+ "0.6": 0.9036458333333334,
432
+ "0.7": 0.8837890625,
433
+ "0.8": 0.8518880208333334,
434
+ "0.9": 0.8037109375
435
+ }
436
+ }
437
+ },
438
+ "summary": {
439
+ "n_seeds": 3,
440
+ "n_pairs": 3,
441
+ "report_band": [
442
+ 0.6,
443
+ 0.7,
444
+ 0.8,
445
+ 0.9
446
+ ],
447
+ "n_dead_mean": 0,
448
+ "n_dead_std": 0.0,
449
+ "n_dense_mean": 12,
450
+ "n_dense_std": 2.516611478423583,
451
+ "l0_mean": 66,
452
+ "l0_std": 1.4188960215604238,
453
+ "recon_r2_mean": 0.9697333333333332,
454
+ "recon_r2_std": 0.0015534906930308038,
455
+ "l1_per_token_mean": 100.54123333333332,
456
+ "l1_per_token_std": 3.709830554531211,
457
+ "babel_mu1_signed_mean": 0.25347430010636646,
458
+ "babel_mu1_signed_std": 0.007525918313292953,
459
+ "n_dup_signed_per_threshold": {
460
+ "0.5": 61.666666666666664,
461
+ "0.6": 0.0,
462
+ "0.7": 0.0,
463
+ "0.8": 0.0,
464
+ "0.9": 0.0
465
+ },
466
+ "n_dup_signed_band": 0.0,
467
+ "n_dup_signed_band_std": 0.0,
468
+ "n_dup_clusters_signed_per_threshold": {
469
+ "0.5": 30.333333333333332,
470
+ "0.6": 0.0,
471
+ "0.7": 0.0,
472
+ "0.8": 0.0,
473
+ "0.9": 0.0
474
+ },
475
+ "n_dup_clusters_signed_band": 0.0,
476
+ "n_dup_clusters_signed_band_std": 0.0,
477
+ "largest_dup_signed_per_threshold": {
478
+ "0.5": 2.6666666666666665,
479
+ "0.6": 1.0,
480
+ "0.7": 1.0,
481
+ "0.8": 1.0,
482
+ "0.9": 1.0
483
+ },
484
+ "largest_dup_signed_band": 1.0,
485
+ "largest_dup_signed_band_std": 0.0,
486
+ "dcs_signed_mean": 0.9010104735692342,
487
+ "dcs_signed_std": 0.004033694346817725,
488
+ "dedup_dcs_signed_per_threshold": {
489
+ "0.5": 0.8934071490657516,
490
+ "0.6": 0.9010104338328043,
491
+ "0.7": 0.9010104338328043,
492
+ "0.8": 0.9010104338328043,
493
+ "0.9": 0.9010104338328043
494
+ },
495
+ "dedup_dcs_signed_band": 0.9010104338328043,
496
+ "dedup_dcs_signed_band_std": 0.0040336830862947705,
497
+ "matched_cos_signed_per_threshold": {
498
+ "0.5": 0.9026136994361877,
499
+ "0.6": 0.9010104338328043,
500
+ "0.7": 0.9010104338328043,
501
+ "0.8": 0.9010104338328043,
502
+ "0.9": 0.9010104338328043
503
+ },
504
+ "matched_cos_signed_band": 0.9010104338328043,
505
+ "matched_cos_signed_band_std": 0.0040336830862947705,
506
+ "unique_frac_signed_per_threshold": {
507
+ "0.5": 0.9898003472222222,
508
+ "0.6": 1.0,
509
+ "0.7": 1.0,
510
+ "0.8": 1.0,
511
+ "0.9": 1.0
512
+ },
513
+ "unique_frac_signed_band": 1.0,
514
+ "unique_frac_signed_band_std": 0.0,
515
+ "matched_frac_signed_per_threshold": {
516
+ "0.5": 0.9127604166666666,
517
+ "0.6": 0.907009548611111,
518
+ "0.7": 0.888671875,
519
+ "0.8": 0.8579644097222222,
520
+ "0.9": 0.8094618055555557
521
+ },
522
+ "matched_frac_signed_band": 0.8657769097222223,
523
+ "matched_frac_signed_band_std": 0.0066624169579947694,
524
+ "babel_mu1_abs_mean": 0.25976213812828064,
525
+ "babel_mu1_abs_std": 0.007369158310030964,
526
+ "n_dup_abs_per_threshold": {
527
+ "0.5": 79.33333333333333,
528
+ "0.6": 10.0,
529
+ "0.7": 5.333333333333333,
530
+ "0.8": 2.0,
531
+ "0.9": 0.0
532
+ },
533
+ "n_dup_abs_band": 4.333333333333333,
534
+ "n_dup_abs_band_std": 0.7637626158259734,
535
+ "n_dup_clusters_abs_per_threshold": {
536
+ "0.5": 38.0,
537
+ "0.6": 5.0,
538
+ "0.7": 2.6666666666666665,
539
+ "0.8": 1.0,
540
+ "0.9": 0.0
541
+ },
542
+ "n_dup_clusters_abs_band": 2.1666666666666665,
543
+ "n_dup_clusters_abs_band_std": 0.3818813079129867,
544
+ "largest_dup_abs_per_threshold": {
545
+ "0.5": 3.6666666666666665,
546
+ "0.6": 2.0,
547
+ "0.7": 2.0,
548
+ "0.8": 2.0,
549
+ "0.9": 1.0
550
+ },
551
+ "largest_dup_abs_band": 1.75,
552
+ "largest_dup_abs_band_std": 0.0,
553
+ "dcs_abs_mean": 0.9014946222305298,
554
+ "dcs_abs_std": 0.0038211313163758432,
555
+ "dedup_dcs_abs_per_threshold": {
556
+ "0.5": 0.8912512103043909,
557
+ "0.6": 0.9002383778408531,
558
+ "0.7": 0.9008646393485833,
559
+ "0.8": 0.9012148914173141,
560
+ "0.9": 0.9014945824941
561
+ },
562
+ "dedup_dcs_abs_band": 0.9009531227752127,
563
+ "dedup_dcs_abs_band_std": 0.00374613458608213,
564
+ "matched_cos_abs_per_threshold": {
565
+ "0.5": 0.9034086664517721,
566
+ "0.6": 0.9017063975334167,
567
+ "0.7": 0.9016475478808085,
568
+ "0.8": 0.9015083511670431,
569
+ "0.9": 0.9014945824941
570
+ },
571
+ "matched_cos_abs_band": 0.9015892197688421,
572
+ "matched_cos_abs_band_std": 0.003804581877868711,
573
+ "unique_frac_abs_per_threshold": {
574
+ "0.5": 0.9865451388888888,
575
+ "0.6": 0.9983723958333334,
576
+ "0.7": 0.9991319444444443,
577
+ "0.8": 0.9996744791666666,
578
+ "0.9": 1.0
579
+ },
580
+ "unique_frac_abs_band": 0.999294704861111,
581
+ "unique_frac_abs_band_std": 6.215516079313338e-05,
582
+ "matched_frac_abs_per_threshold": {
583
+ "0.5": 0.9112413194444443,
584
+ "0.6": 0.9055989583333334,
585
+ "0.7": 0.8885633680555555,
586
+ "0.8": 0.8582899305555557,
587
+ "0.9": 0.8092447916666666
588
+ },
589
+ "matched_frac_abs_band": 0.8654242621527778,
590
+ "matched_frac_abs_band_std": 0.006400817663006407,
591
+ "alive_source": "stats",
592
+ "frac_denominator": "d_sae",
593
+ "d_sae": 3072
594
+ }
595
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.08,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.112,
3
+ "r2_score": 0.971,
4
+ "l0_per_token_0": 0.0219,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0013,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 103.8579
9
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.08,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1153,
3
+ "r2_score": 0.9702,
4
+ "l0_per_token_0": 0.0219,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.001,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 101.2308
9
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 1.1,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 5,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 0,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
vision/dinov2-b_layer10/consdl/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.118,
3
+ "r2_score": 0.968,
4
+ "l0_per_token_0": 0.0211,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.001,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 96.535
9
+ }
vision/dinov2-b_layer10/consdl/x8/l0_64/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 61.49767833981842,
20
+ "recon_r2": 0.9570094472047507,
21
+ "l1_per_token": 62.24012274897282
22
+ },
23
+ "seed1": {
24
+ "l0": 61.51243968871595,
25
+ "recon_r2": 0.9569624626355632,
26
+ "l1_per_token": 61.303796587848595
27
+ },
28
+ "seed2": {
29
+ "l0": 60.979823605706876,
30
+ "recon_r2": 0.9570221119832084,
31
+ "l1_per_token": 61.43573532514122
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.8091336131732522,
38
+ "mcc_abs": 0.809143052980791,
39
+ "act_iou": 0.5349657871026813,
40
+ "act_iou_n_pairs": 6144,
41
+ "input_r2": 0.9967531489632794,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8573026456548342,
44
+ "0.7": 0.881288445742012,
45
+ "0.8": 0.9139149550931535,
46
+ "0.9": 0.9527607266618047
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8571283782181095,
50
+ "0.7": 0.8815317068035863,
51
+ "0.8": 0.9139577148316728,
52
+ "0.9": 0.952745001238072
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.811658921317132,
58
+ "mcc_abs": 0.8116677447318766,
59
+ "act_iou": 0.5387286135389245,
60
+ "act_iou_n_pairs": 6144,
61
+ "input_r2": 0.9967930258734372,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8601261105655228,
64
+ "0.7": 0.8816813929345667,
65
+ "0.8": 0.9163874123933098,
66
+ "0.9": 0.9536953979153595
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.859874159140255,
70
+ "0.7": 0.8811639953940094,
71
+ "0.8": 0.9161314175167206,
72
+ "0.9": 0.9537596005155964
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.8126490489417808,
78
+ "mcc_abs": 0.8126618279513241,
79
+ "act_iou": 0.5392205481575283,
80
+ "act_iou_n_pairs": 6144,
81
+ "input_r2": 0.9968035572401793,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.8597588708602656,
84
+ "0.7": 0.8840962278255189,
85
+ "0.8": 0.9184536996185644,
86
+ "0.9": 0.9548546368779008
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.859722841080943,
90
+ "0.7": 0.8840157745625136,
91
+ "0.8": 0.9183664355361395,
92
+ "0.9": 0.9548853054491077
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 61.32998054474709,
106
+ "l0_std": 0.30333461032962217,
107
+ "recon_r2_mean": 0.9569980072745073,
108
+ "recon_r2_std": 3.142714068450393e-05,
109
+ "l1_per_token_mean": 61.65988488732088,
110
+ "l1_per_token_std": 0.5068125291174315,
111
+ "input_r2_mean": 0.9967832440256319,
112
+ "input_r2_std": 2.658969734444384e-05,
113
+ "mcc_signed_mean": 0.8111471944773884,
114
+ "mcc_signed_std": 0.0018127245873403347,
115
+ "mcc_abs_mean": 0.8111575418879973,
116
+ "mcc_abs_std": 0.001814021645333788,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8590625423602075,
119
+ "0.7": 0.8823553555006992,
120
+ "0.8": 0.9162520223683425,
121
+ "0.9": 0.9537702538183549
122
+ },
123
+ "matched_mcc_signed_band": 0.902860043511901,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.858908459479769,
126
+ "0.7": 0.8822371589200363,
127
+ "0.8": 0.9161518559615108,
128
+ "0.9": 0.9537966357342587
129
+ },
130
+ "matched_mcc_abs_band": 0.9027735275238937,
131
+ "act_iou_mean": 0.537638316266378,
132
+ "act_iou_std": 0.0023275113351716404,
133
+ "act_iou_n_pairs": 6144.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }