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
4d86c3c
·
verified ·
1 Parent(s): 7a5b70c

add samelayer_reduce llama32-3b-grpo

Browse files
Files changed (4) hide show
  1. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/activations/activations.pt +3 -0
  2. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/checkpoints/final.pt +3 -0
  3. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/metrics/training_metrics.json +80 -0
  4. v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/run_meta.json +34 -0
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/activations/activations.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ce6a49d8ce4dbc9f8c064d23305d327548664177437b836d1732fd03ade3a95
3
+ size 1501941701
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/checkpoints/final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:503c52a98becdb0e546155640654d9db3a4ab6d1caf97d65059943c5524dd35e
3
+ size 3624487706
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/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
+ 0.04863162126807497,
10
+ 0.028767385191115036,
11
+ 0.02664078392906107,
12
+ 0.02561670167603238
13
+ ],
14
+ "val_loss": [
15
+ 0.030831219794238425,
16
+ 0.028230151346840784,
17
+ 0.0271887624568028,
18
+ 0.026992113831701702
19
+ ],
20
+ "train_fve_base": [
21
+ 0.6179791546605584,
22
+ 0.7524986343639883,
23
+ 0.7778358430592414,
24
+ 0.790156301287135
25
+ ],
26
+ "train_fve_aligned": [
27
+ 0.6104102654548952,
28
+ 0.7452379631077555,
29
+ 0.770189265827651,
30
+ 0.7819857032620552
31
+ ],
32
+ "val_fve_base": [
33
+ 0.729199195407448,
34
+ 0.7597596046188115,
35
+ 0.7722388240679396,
36
+ 0.7752108305536639
37
+ ],
38
+ "val_fve_aligned": [
39
+ 0.722105749926642,
40
+ 0.7537020919210624,
41
+ 0.7663287392461487,
42
+ 0.7692189825142866
43
+ ],
44
+ "dead_neurons": [
45
+ 0.9139503546089578,
46
+ 0.9561741884452032,
47
+ 0.955374258238185,
48
+ 0.9546477757764016
49
+ ],
50
+ "l0_base": [
51
+ 221.1378064798599,
52
+ 210.38729568009342,
53
+ 212.65694322825453,
54
+ 213.7158676298891
55
+ ],
56
+ "l0_aligned": [
57
+ 217.28639448336253,
58
+ 208.17113616462348,
59
+ 211.52583187390542,
60
+ 212.69061952714537
61
+ ],
62
+ "self_recon": [
63
+ 0.026460523027176407,
64
+ 0.01723270824584638,
65
+ 0.015504534835821171,
66
+ 0.014676261198491084
67
+ ],
68
+ "cross_recon": [
69
+ 0.026530061130805003,
70
+ 0.017257546096979284,
71
+ 0.015529198475264695,
72
+ 0.014701342010099639
73
+ ],
74
+ "sparsity": [
75
+ 0.011559073629631931,
76
+ 0.004631658412121406,
77
+ 0.004924569560581317,
78
+ 0.005059903606628891
79
+ ]
80
+ }
v3-samelayer-seed-99/samelayer_reduce/llama32-3b-grpo/L12-14/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-grpo",
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-grpo/activations/activations.pt",
7
+ "base_activations_path": "/home/igarg/research/interp-alignment/interp_utils/crosscoder/results-multi-v1/v1/assembled_activations/llama32-3b-grpo/activations/activations.pt",
8
+ "source_layers": [
9
+ 12,
10
+ 13,
11
+ 14
12
+ ],
13
+ "base_source_layers": [
14
+ 12,
15
+ 13,
16
+ 14
17
+ ],
18
+ "layers": [
19
+ 12,
20
+ 13,
21
+ 14
22
+ ],
23
+ "base_layers": [
24
+ 12,
25
+ 13,
26
+ 14
27
+ ],
28
+ "aligned_layers": [
29
+ 12,
30
+ 13,
31
+ 14
32
+ ],
33
+ "layer_reduction": "mean"
34
+ }