Hugging Face's logo

Datasets:
MInAlA
/
crosscoder-multilayer-split-activations

Tags:
mechanistic-interpretability
crosscoder
activations
sparse-autoencoder
License:
Dataset card Files
xet
 Community
1
ishangarg183 commited on
Commit
caaa03d
·
verified ·
1 Parent(s): ac3f3f6

add samelayer_separate llama32-3b-grpo

Browse files
Files changed (12) hide show
  1. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/activations/activations.pt +3 -0
  2. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/checkpoints/final.pt +3 -0
  3. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/metrics/training_metrics.json +80 -0
  4. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/run_meta.json +27 -0
  5. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/activations/activations.pt +3 -0
  6. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/checkpoints/final.pt +3 -0
  7. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/metrics/training_metrics.json +80 -0
  8. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/run_meta.json +27 -0
  9. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/activations/activations.pt +3 -0
  10. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/checkpoints/final.pt +3 -0
  11. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/metrics/training_metrics.json +80 -0
  12. v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/run_meta.json +27 -0
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/activations/activations.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:11218523d909b4fff41c01a116fef25c28f6d5704006255e3dd523fee8c6c6aa
3
+ size 1501941701
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/checkpoints/final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2045ba6b2d3ad5942cacd3568f803ee24c4cfc4cf9e91afba3e47babc5a6a637
3
+ size 3624487706
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/metrics/training_metrics.json ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epochs": [
3
+ 1,
4
+ 2,
5
+ 3,
6
+ 4
7
+ ],
8
+ "train_loss": [
9
+ 0.052543846299894635,
10
+ 0.037991779263229726,
11
+ 0.03663025246644953,
12
+ 0.03587210086928393
13
+ ],
14
+ "val_loss": [
15
+ 0.03922261709240094,
16
+ 0.03764664721114473,
17
+ 0.036965264992408105,
18
+ 0.036792815024159965
19
+ ],
20
+ "train_fve_base": [
21
+ 0.5381988162752208,
22
+ 0.6582887355375207,
23
+ 0.6841748292447528,
24
+ 0.6982198739288428
25
+ ],
26
+ "train_fve_aligned": [
27
+ 0.44409866010801735,
28
+ 0.5548158268168595,
29
+ 0.5775612378357031,
30
+ 0.5906626017415169
31
+ ],
32
+ "val_fve_base": [
33
+ 0.638550155137846,
34
+ 0.6641819408426735,
35
+ 0.6757678891975842,
36
+ 0.6792140752857269
37
+ ],
38
+ "val_fve_aligned": [
39
+ 0.5370016088660475,
40
+ 0.5644941517195776,
41
+ 0.5787024076696465,
42
+ 0.5831787333438534
43
+ ],
44
+ "dead_neurons": [
45
+ 0.9054380945482938,
46
+ 0.9572945319972954,
47
+ 0.9569545930335206,
48
+ 0.9564564310175176
49
+ ],
50
+ "l0_base": [
51
+ 201.5541812609457,
52
+ 181.35818009340338,
53
+ 188.07942936368943,
54
+ 192.72276342673672
55
+ ],
56
+ "l0_aligned": [
57
+ 184.8096176298891,
58
+ 168.25193374197315,
59
+ 183.46282107413893,
60
+ 190.79994892002335
61
+ ],
62
+ "self_recon": [
63
+ 0.028374901272995877,
64
+ 0.022119998246848897,
65
+ 0.020805126549735865,
66
+ 0.02006713669340442
67
+ ],
68
+ "cross_recon": [
69
+ 0.03161178173043891,
70
+ 0.02647448546849561,
71
+ 0.025468757095536292,
72
+ 0.024900639480794306
73
+ ],
74
+ "sparsity": [
75
+ 0.011524232116381606,
76
+ 0.005281986563504003,
77
+ 0.005637622840421995,
78
+ 0.005844708075948055
79
+ ]
80
+ }
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L10__aligned_L12/run_meta.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_model": "meta-llama/Llama-3.2-3B-Instruct",
3
+ "aligned_run_id": "llama32-3b-grpo",
4
+ "position": "last_prompt",
5
+ "multi_type": "separate",
6
+ "source_activations": "/home/ubuntu/interp-alignment-crosscoder/v1/assembled_activations/llama32-3b-grpo/activations/activations.pt",
7
+ "base_activations_path": "/home/ubuntu/interp-alignment-crosscoder/v1/base_activations/llama32-3b-union/activations/base_activations.pt",
8
+ "source_layers": [
9
+ 12,
10
+ 13,
11
+ 14
12
+ ],
13
+ "base_source_layers": [
14
+ 10,
15
+ 11,
16
+ 12,
17
+ 13,
18
+ 14,
19
+ 23,
20
+ 24,
21
+ 25,
22
+ 26
23
+ ],
24
+ "layer": 12,
25
+ "base_layer": 10,
26
+ "aligned_layer": 12
27
+ }
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/activations/activations.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3ca7e810884aaf3a11e297acc42174b2ab1fd261163bc101aa98e5029a29d487
3
+ size 1501941701
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/checkpoints/final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2aff04b40d34c1cc8f99aed7ba3abbe6d0444ec4eef52e31d92fd44fbaa52aba
3
+ size 3624487706
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/metrics/training_metrics.json ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epochs": [
3
+ 1,
4
+ 2,
5
+ 3,
6
+ 4
7
+ ],
8
+ "train_loss": [
9
+ 0.0602064708772975,
10
+ 0.04275895518148008,
11
+ 0.04116938315153609,
12
+ 0.040305447561902324
13
+ ],
14
+ "val_loss": [
15
+ 0.04419337019948435,
16
+ 0.042579951710725955,
17
+ 0.04185814205654629,
18
+ 0.04168874832816149
19
+ ],
20
+ "train_fve_base": [
21
+ 0.586389712054202,
22
+ 0.7044924112418487,
23
+ 0.7267465177364538,
24
+ 0.7388025726829436
25
+ ],
26
+ "train_fve_aligned": [
27
+ 0.45764722259714113,
28
+ 0.5645684315945346,
29
+ 0.584718160579167,
30
+ 0.5960258754873861
31
+ ],
32
+ "val_fve_base": [
33
+ 0.6881447668475006,
34
+ 0.7076546992931066,
35
+ 0.7165167462763362,
36
+ 0.7197755668799914
37
+ ],
38
+ "val_fve_aligned": [
39
+ 0.5484183992390858,
40
+ 0.5709807130054654,
41
+ 0.5820607586680907,
42
+ 0.5855631687878314
43
+ ],
44
+ "dead_neurons": [
45
+ 0.8989692701280152,
46
+ 0.9502163075663093,
47
+ 0.9503997329446355,
48
+ 0.9500647589468636
49
+ ],
50
+ "l0_base": [
51
+ 201.39422796263864,
52
+ 176.48201255107998,
53
+ 183.49802977232926,
54
+ 188.5417578809107
55
+ ],
56
+ "l0_aligned": [
57
+ 177.46845811441915,
58
+ 152.24708114419147,
59
+ 166.35247008172797,
60
+ 175.5306844716871
61
+ ],
62
+ "self_recon": [
63
+ 0.03308812573756555,
64
+ 0.025460599506816457,
65
+ 0.02401948203612717,
66
+ 0.023224963865830617
67
+ ],
68
+ "cross_recon": [
69
+ 0.03631520044804489,
70
+ 0.029941234835496926,
71
+ 0.028750915056387025,
72
+ 0.028084338795430677
73
+ ],
74
+ "sparsity": [
75
+ 0.012592264657140962,
76
+ 0.005321861503529431,
77
+ 0.005649534824494026,
78
+ 0.005846747940964018
79
+ ]
80
+ }
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L11__aligned_L13/run_meta.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_model": "meta-llama/Llama-3.2-3B-Instruct",
3
+ "aligned_run_id": "llama32-3b-grpo",
4
+ "position": "last_prompt",
5
+ "multi_type": "separate",
6
+ "source_activations": "/home/ubuntu/interp-alignment-crosscoder/v1/assembled_activations/llama32-3b-grpo/activations/activations.pt",
7
+ "base_activations_path": "/home/ubuntu/interp-alignment-crosscoder/v1/base_activations/llama32-3b-union/activations/base_activations.pt",
8
+ "source_layers": [
9
+ 12,
10
+ 13,
11
+ 14
12
+ ],
13
+ "base_source_layers": [
14
+ 10,
15
+ 11,
16
+ 12,
17
+ 13,
18
+ 14,
19
+ 23,
20
+ 24,
21
+ 25,
22
+ 26
23
+ ],
24
+ "layer": 13,
25
+ "base_layer": 11,
26
+ "aligned_layer": 13
27
+ }
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/activations/activations.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dbc81af395f85a014eb8279b4d2ca05eab253827d69218837ba3b724042b83cd
3
+ size 1501941701
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/checkpoints/final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f5b833c21fd74335a5410b31fe94df41e4ec439a0bff07637479362db564f396
3
+ size 3624487706
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/metrics/training_metrics.json ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epochs": [
3
+ 1,
4
+ 2,
5
+ 3,
6
+ 4
7
+ ],
8
+ "train_loss": [
9
+ 0.06937128079047136,
10
+ 0.04941633406328445,
11
+ 0.04734274588842804,
12
+ 0.04617825649587369
13
+ ],
14
+ "val_loss": [
15
+ 0.051233707395679666,
16
+ 0.04895065313312396,
17
+ 0.047810977334595474,
18
+ 0.04751443937030762
19
+ ],
20
+ "train_fve_base": [
21
+ 0.5764109695507501,
22
+ 0.6988560761962798,
23
+ 0.7226747293536314,
24
+ 0.7359284788143962
25
+ ],
26
+ "train_fve_aligned": [
27
+ 0.4907508262471713,
28
+ 0.6033714425473344,
29
+ 0.6247510264230206,
30
+ 0.6374478014811256
31
+ ],
32
+ "val_fve_base": [
33
+ 0.6789273036087995,
34
+ 0.7047899967712882,
35
+ 0.7161360944752918,
36
+ 0.7195423192378738
37
+ ],
38
+ "val_fve_aligned": [
39
+ 0.5864810846862992,
40
+ 0.6113509326705133,
41
+ 0.6254426749588932,
42
+ 0.6299409894418966
43
+ ],
44
+ "dead_neurons": [
45
+ 0.9116211133746466,
46
+ 0.9590461073183415,
47
+ 0.9585443808541129,
48
+ 0.9579262856635447
49
+ ],
50
+ "l0_base": [
51
+ 212.0691951255108,
52
+ 194.87096833041448,
53
+ 197.7655976357268,
54
+ 201.4207348219498
55
+ ],
56
+ "l0_aligned": [
57
+ 203.01209500875657,
58
+ 189.49857705779334,
59
+ 199.9386492994746,
60
+ 205.1522913018097
61
+ ],
62
+ "self_recon": [
63
+ 0.03811953175036586,
64
+ 0.028749232486508007,
65
+ 0.026945049644508213,
66
+ 0.025905121369010096
67
+ ],
68
+ "cross_recon": [
69
+ 0.042258919756565955,
70
+ 0.03368262950520173,
71
+ 0.03168530664768263,
72
+ 0.030743483610568657
73
+ ],
74
+ "sparsity": [
75
+ 0.01434818075119501,
76
+ 0.007194049469909157,
77
+ 0.007723573367483805,
78
+ 0.007975741408893624
79
+ ]
80
+ }
v3-samelayer-seed-99/samelayer_separate/llama32-3b-grpo/base_L12__aligned_L14/run_meta.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_model": "meta-llama/Llama-3.2-3B-Instruct",
3
+ "aligned_run_id": "llama32-3b-grpo",
4
+ "position": "last_prompt",
5
+ "multi_type": "separate",
6
+ "source_activations": "/home/ubuntu/interp-alignment-crosscoder/v1/assembled_activations/llama32-3b-grpo/activations/activations.pt",
7
+ "base_activations_path": "/home/ubuntu/interp-alignment-crosscoder/v1/base_activations/llama32-3b-union/activations/base_activations.pt",
8
+ "source_layers": [
9
+ 12,
10
+ 13,
11
+ 14
12
+ ],
13
+ "base_source_layers": [
14
+ 10,
15
+ 11,
16
+ 12,
17
+ 13,
18
+ 14,
19
+ 23,
20
+ 24,
21
+ 25,
22
+ 26
23
+ ],
24
+ "layer": 14,
25
+ "base_layer": 12,
26
+ "aligned_layer": 14
27
+ }