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
9c05935
·
verified ·
1 Parent(s): 4d86c3c

add samelayer_reduce llama32-3b-kto

Browse files
Files changed (4) hide show
  1. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/activations/activations.pt +3 -0
  2. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/checkpoints/final.pt +3 -0
  3. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/metrics/training_metrics.json +80 -0
  4. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/run_meta.json +34 -0
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/activations/activations.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ebc3b77a473e7846d9727ec9dfc97377a27cfe09c4e63a088d0852cc2a36a908
3
+ size 1501941701
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/checkpoints/final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a9602b4e273c22695d8576e326e788d3e13e356b37584c0fcac1b656927633e7
3
+ size 3624487706
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/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
+ 0.4299773222789107,
10
+ 0.3012607290740606,
11
+ 0.2821249345625205,
12
+ 0.272205770250586
13
+ ],
14
+ "val_loss": [
15
+ 0.31692506224697176,
16
+ 0.2944472590510134,
17
+ 0.2854639206725265,
18
+ 0.283354187979124
19
+ ],
20
+ "train_fve_base": [
21
+ 0.5540539581596121,
22
+ 0.6849707413749673,
23
+ 0.7119429785816974,
24
+ 0.7264657600957333
25
+ ],
26
+ "train_fve_aligned": [
27
+ 0.6270287533775103,
28
+ 0.7443730133497012,
29
+ 0.7655923712065675,
30
+ 0.776828084086648
31
+ ],
32
+ "val_fve_base": [
33
+ 0.661541308095942,
34
+ 0.6931266962545705,
35
+ 0.7067132273893706,
36
+ 0.7101863232582651
37
+ ],
38
+ "val_fve_aligned": [
39
+ 0.7286595112366202,
40
+ 0.7538824733639262,
41
+ 0.7650630059042526,
42
+ 0.7684432514675
43
+ ],
44
+ "dead_neurons": [
45
+ 0.9330430873569103,
46
+ 0.9680153655852759,
47
+ 0.9677400830817793,
48
+ 0.9673738492488583
49
+ ],
50
+ "l0_base": [
51
+ 231.38246132516053,
52
+ 223.87428852889667,
53
+ 222.63096541155866,
54
+ 222.34831071220083
55
+ ],
56
+ "l0_aligned": [
57
+ 226.31003721541157,
58
+ 222.3933523058961,
59
+ 221.79250583771162,
60
+ 221.16051882661998
61
+ ],
62
+ "self_recon": [
63
+ 0.26704640623010256,
64
+ 0.18571901624131745,
65
+ 0.1700535206786388,
66
+ 0.16168809925699124
67
+ ],
68
+ "cross_recon": [
69
+ 0.28840285629225415,
70
+ 0.2122281804828705,
71
+ 0.19725685749774918,
72
+ 0.18884100698443512
73
+ ],
74
+ "sparsity": [
75
+ 0.04756977025286816,
76
+ 0.030650439308450645,
77
+ 0.033168668483024584,
78
+ 0.034981266351311234
79
+ ]
80
+ }
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-kto/L23-25/run_meta.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_model": "meta-llama/Llama-3.2-3B-Instruct",
3
+ "aligned_run_id": "llama32-3b-kto",
4
+ "position": "last_prompt",
5
+ "multi_type": "reduce",
6
+ "source_activations": "/home/igarg/research/interp-alignment/interp_utils/crosscoder/results-multi-v1/v1/assembled_activations/llama32-3b-kto/activations/activations.pt",
7
+ "base_activations_path": "/home/igarg/research/interp-alignment/interp_utils/crosscoder/results-multi-v1/v1/assembled_activations/llama32-3b-kto/activations/activations.pt",
8
+ "source_layers": [
9
+ 23,
10
+ 24,
11
+ 25
12
+ ],
13
+ "base_source_layers": [
14
+ 23,
15
+ 24,
16
+ 25
17
+ ],
18
+ "layers": [
19
+ 23,
20
+ 24,
21
+ 25
22
+ ],
23
+ "base_layers": [
24
+ 23,
25
+ 24,
26
+ 25
27
+ ],
28
+ "aligned_layers": [
29
+ 23,
30
+ 24,
31
+ 25
32
+ ],
33
+ "layer_reduction": "mean"
34
+ }