Hugging Face's logo

hematteo
/
parameter-trajectory-crosscoders

crosscoder
sparse-autoencoder
mech-interp
parameter-trajectory
pythia
olmo
Model card Files
xet
 Community
Matteo He commited on
Commit
0388f08
·
verified ·
1 Parent(s): 1373830

Add files using upload-large-folder tool

Browse files
Files changed (20) hide show
  1. derived/rates/wu-d8192-multiseed/decoder_norms_all_seeds.npy +3 -0
  2. derived/rates/wu-d8192-multiseed/firing_rates_all_seeds.npy +3 -0
  3. derived/rates/wu-d8192-persnap/train.log +965 -0
  4. evaluation/eval-corpus/README.md +8 -0
  5. evaluation/eval-corpus/eval_tokens.pt +3 -0
  6. olmo-2-7b/W_U/cross-snapshot-32/d32768/seed0.config.json +118 -0
  7. olmo-2-7b/W_U/cross-snapshot-32/d32768/seed0.md +12 -0
  8. pythia-160m/W_U/lambda-sweep/d8192/lam0p40_seed0.metrics_recomputed.json +24 -0
  9. pythia-160m/W_U/lambda-sweep/d8192/lam1p00_seed0.metrics_recomputed.json +24 -0
  10. pythia-160m/W_U/lambda-sweep/d8192/lam1p20_seed0.metrics_recomputed.json +24 -0
  11. pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed0.metrics_recomputed.json +24 -0
  12. pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed1.metrics_recomputed.json +24 -0
  13. pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed2.metrics_recomputed.json +24 -0
  14. pythia-160m/W_U/lambda-sweep/d8192/lam1p80_seed0.metrics_recomputed.json +24 -0
  15. pythia-1b/W_U/cross-snapshot-32-matched-window/d24576/seed0.config.json +118 -0
  16. pythia-1b/W_U/cross-snapshot-32-matched-window/d24576/seed0.md +8 -0
  17. pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0-sparse.config.json +118 -0
  18. pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0-sparse.md +8 -0
  19. pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0.config.json +118 -0
  20. pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0.md +6 -0
derived/rates/wu-d8192-multiseed/decoder_norms_all_seeds.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:143fb58a869fbc1d7360984991ac943f138756109ddf1c9262ddcc63bd9e4acd
3
+ size 5243008
derived/rates/wu-d8192-multiseed/firing_rates_all_seeds.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0aca81e20992077e8ee9bdb51e158427126616d12dcf2cd43758314dc52ce5af
3
+ size 5243008
derived/rates/wu-d8192-persnap/train.log ADDED
@@ -0,0 +1,965 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ git_hash=unknown torch=2.7.1+cu118 device=cuda
2
+ output_dir=/home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218
3
+ cache_dir=/home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
4
+ [step 0] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
5
+ [step 0] preprocess center_scale: mean-shift=0.003 scale=0.0228
6
+ [epoch 3 step 100] loss 760.4553 l_rec 760.3167 l_s 1.1670 l_aux 0.0000 l1x 0.40 dead 0/8192
7
+ [epoch 7 step 200] loss 695.5469 l_rec 690.8574 l_s 19.6377 l_aux 0.0000 l1x 0.80 dead 0/8192
8
+ [epoch 11 step 300] loss 468.2347 l_rec 269.4035 l_s 662.7706 l_aux 0.0000 l1x 1.00 dead 0/8192
9
+ [epoch 15 step 400] loss 458.0808 l_rec 254.1990 l_s 679.6062 l_aux 0.0000 l1x 1.00 dead 0/8192
10
+ [epoch 19 step 500] loss 454.0783 l_rec 246.9091 l_s 690.5642 l_aux 0.0000 l1x 1.00 dead 0/8192
11
+ [epoch 23 step 600] loss 450.4685 l_rec 242.1719 l_s 694.3221 l_aux 0.0000 l1x 1.00 dead 0/8192
12
+ [epoch 27 step 700] loss 447.8918 l_rec 238.7377 l_s 697.1802 l_aux 0.0000 l1x 1.00 dead 0/8192
13
+ [epoch 31 step 800] loss 445.7771 l_rec 235.5080 l_s 700.8969 l_aux 0.0000 l1x 1.00 dead 0/8192
14
+ [epoch 35 step 900] loss 443.7163 l_rec 232.9742 l_s 702.4738 l_aux 0.0000 l1x 1.00 dead 0/8192
15
+ [epoch 39 step 1000] loss 440.5496 l_rec 230.1541 l_s 701.3182 l_aux 0.0000 l1x 1.00 dead 0/8192
16
+ [epoch 43 step 1100] loss 438.2469 l_rec 227.8399 l_s 701.3567 l_aux 0.0000 l1x 1.00 dead 0/8192
17
+ [epoch 47 step 1200] loss 436.6033 l_rec 225.4635 l_s 703.7993 l_aux 0.0000 l1x 1.00 dead 0/8192
18
+ [epoch 51 step 1300] loss 433.7432 l_rec 223.1875 l_s 701.8521 l_aux 0.0000 l1x 1.00 dead 0/8192
19
+ [epoch 55 step 1400] loss 432.3996 l_rec 221.5412 l_s 702.8613 l_aux 0.0000 l1x 1.00 dead 0/8192
20
+ [epoch 59 step 1500] loss 429.9479 l_rec 219.4557 l_s 701.6406 l_aux 0.0000 l1x 1.00 dead 0/8192
21
+ [epoch 63 step 1600] loss 428.2051 l_rec 218.0668 l_s 700.4610 l_aux 0.0000 l1x 1.00 dead 0/8192
22
+ [epoch 67 step 1700] loss 426.9931 l_rec 216.3646 l_s 702.0951 l_aux 0.0000 l1x 1.00 dead 0/8192
23
+ [epoch 71 step 1800] loss 426.0205 l_rec 215.0011 l_s 703.3981 l_aux 0.0000 l1x 1.00 dead 0/8192
24
+ [epoch 75 step 1900] loss 422.8461 l_rec 213.3312 l_s 698.3829 l_aux 0.0000 l1x 1.00 dead 0/8192
25
+ [epoch 79 step 2000] loss 422.1132 l_rec 211.9946 l_s 700.3953 l_aux 0.0000 l1x 1.00 dead 0/8192
26
+ [epoch 83 step 2100] loss 420.5233 l_rec 210.5915 l_s 699.7727 l_aux 0.0000 l1x 1.00 dead 0/8192
27
+ [epoch 87 step 2200] loss 418.2918 l_rec 209.5669 l_s 695.7496 l_aux 0.0000 l1x 1.00 dead 0/8192
28
+ [epoch 91 step 2300] loss 418.2252 l_rec 208.7844 l_s 698.1360 l_aux 0.0000 l1x 1.00 dead 0/8192
29
+ [epoch 95 step 2400] loss 417.1713 l_rec 208.2415 l_s 696.4328 l_aux 0.0000 l1x 1.00 dead 0/8192
30
+ [epoch 99 step 2500] loss 416.4102 l_rec 208.0490 l_s 694.5376 l_aux 0.0000 l1x 1.00 dead 0/8192
31
+ [step 0] EV=0.7289 L0=1150.1 (40.3s)
32
+ [step 0] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step0.pt
33
+ [step 1] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
34
+ [step 1] preprocess center_scale: mean-shift=0.003 scale=0.0228
35
+ [epoch 3 step 100] loss 760.4553 l_rec 760.3167 l_s 1.1670 l_aux 0.0000 l1x 0.40 dead 0/8192
36
+ [epoch 7 step 200] loss 695.5469 l_rec 690.8574 l_s 19.6377 l_aux 0.0000 l1x 0.80 dead 0/8192
37
+ [epoch 11 step 300] loss 468.2347 l_rec 269.4035 l_s 662.7706 l_aux 0.0000 l1x 1.00 dead 0/8192
38
+ [epoch 15 step 400] loss 458.0808 l_rec 254.1990 l_s 679.6062 l_aux 0.0000 l1x 1.00 dead 0/8192
39
+ [epoch 19 step 500] loss 454.0783 l_rec 246.9091 l_s 690.5642 l_aux 0.0000 l1x 1.00 dead 0/8192
40
+ [epoch 23 step 600] loss 450.4685 l_rec 242.1719 l_s 694.3221 l_aux 0.0000 l1x 1.00 dead 0/8192
41
+ [epoch 27 step 700] loss 447.8918 l_rec 238.7377 l_s 697.1802 l_aux 0.0000 l1x 1.00 dead 0/8192
42
+ [epoch 31 step 800] loss 445.7771 l_rec 235.5080 l_s 700.8969 l_aux 0.0000 l1x 1.00 dead 0/8192
43
+ [epoch 35 step 900] loss 443.7163 l_rec 232.9742 l_s 702.4738 l_aux 0.0000 l1x 1.00 dead 0/8192
44
+ [epoch 39 step 1000] loss 440.5496 l_rec 230.1541 l_s 701.3182 l_aux 0.0000 l1x 1.00 dead 0/8192
45
+ [epoch 43 step 1100] loss 438.2469 l_rec 227.8399 l_s 701.3567 l_aux 0.0000 l1x 1.00 dead 0/8192
46
+ [epoch 47 step 1200] loss 436.6033 l_rec 225.4635 l_s 703.7993 l_aux 0.0000 l1x 1.00 dead 0/8192
47
+ [epoch 51 step 1300] loss 433.7432 l_rec 223.1875 l_s 701.8521 l_aux 0.0000 l1x 1.00 dead 0/8192
48
+ [epoch 55 step 1400] loss 432.3996 l_rec 221.5412 l_s 702.8613 l_aux 0.0000 l1x 1.00 dead 0/8192
49
+ [epoch 59 step 1500] loss 429.9479 l_rec 219.4557 l_s 701.6406 l_aux 0.0000 l1x 1.00 dead 0/8192
50
+ [epoch 63 step 1600] loss 428.2051 l_rec 218.0668 l_s 700.4610 l_aux 0.0000 l1x 1.00 dead 0/8192
51
+ [epoch 67 step 1700] loss 426.9931 l_rec 216.3646 l_s 702.0951 l_aux 0.0000 l1x 1.00 dead 0/8192
52
+ [epoch 71 step 1800] loss 426.0205 l_rec 215.0011 l_s 703.3981 l_aux 0.0000 l1x 1.00 dead 0/8192
53
+ [epoch 75 step 1900] loss 422.8461 l_rec 213.3312 l_s 698.3829 l_aux 0.0000 l1x 1.00 dead 0/8192
54
+ [epoch 79 step 2000] loss 422.1132 l_rec 211.9946 l_s 700.3953 l_aux 0.0000 l1x 1.00 dead 0/8192
55
+ [epoch 83 step 2100] loss 420.5233 l_rec 210.5915 l_s 699.7727 l_aux 0.0000 l1x 1.00 dead 0/8192
56
+ [epoch 87 step 2200] loss 418.2918 l_rec 209.5669 l_s 695.7496 l_aux 0.0000 l1x 1.00 dead 0/8192
57
+ [epoch 91 step 2300] loss 418.2252 l_rec 208.7844 l_s 698.1360 l_aux 0.0000 l1x 1.00 dead 0/8192
58
+ [epoch 95 step 2400] loss 417.1713 l_rec 208.2415 l_s 696.4328 l_aux 0.0000 l1x 1.00 dead 0/8192
59
+ [epoch 99 step 2500] loss 416.4102 l_rec 208.0490 l_s 694.5376 l_aux 0.0000 l1x 1.00 dead 0/8192
60
+ [step 1] EV=0.7289 L0=1150.1 (38.5s)
61
+ [step 1] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step1.pt
62
+ [step 2] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
63
+ [step 2] preprocess center_scale: mean-shift=0.003 scale=0.0228
64
+ [epoch 3 step 100] loss 760.4552 l_rec 760.3166 l_s 1.1670 l_aux 0.0000 l1x 0.40 dead 0/8192
65
+ [epoch 7 step 200] loss 695.5470 l_rec 690.8575 l_s 19.6378 l_aux 0.0000 l1x 0.80 dead 0/8192
66
+ [epoch 11 step 300] loss 468.2342 l_rec 269.4073 l_s 662.7563 l_aux 0.0000 l1x 1.00 dead 0/8192
67
+ [epoch 15 step 400] loss 458.0830 l_rec 254.1963 l_s 679.6224 l_aux 0.0000 l1x 1.00 dead 0/8192
68
+ [epoch 19 step 500] loss 454.0785 l_rec 246.9087 l_s 690.5660 l_aux 0.0000 l1x 1.00 dead 0/8192
69
+ [epoch 23 step 600] loss 450.4722 l_rec 242.1616 l_s 694.3686 l_aux 0.0000 l1x 1.00 dead 0/8192
70
+ [epoch 27 step 700] loss 447.9016 l_rec 238.7309 l_s 697.2357 l_aux 0.0000 l1x 1.00 dead 0/8192
71
+ [epoch 31 step 800] loss 445.7861 l_rec 235.5028 l_s 700.9445 l_aux 0.0000 l1x 1.00 dead 0/8192
72
+ [epoch 35 step 900] loss 443.7143 l_rec 232.9831 l_s 702.4373 l_aux 0.0000 l1x 1.00 dead 0/8192
73
+ [epoch 39 step 1000] loss 440.5602 l_rec 230.1415 l_s 701.3955 l_aux 0.0000 l1x 1.00 dead 0/8192
74
+ [epoch 43 step 1100] loss 438.2501 l_rec 227.8353 l_s 701.3827 l_aux 0.0000 l1x 1.00 dead 0/8192
75
+ [epoch 47 step 1200] loss 436.6097 l_rec 225.4426 l_s 703.8905 l_aux 0.0000 l1x 1.00 dead 0/8192
76
+ [epoch 51 step 1300] loss 433.7414 l_rec 223.1875 l_s 701.8463 l_aux 0.0000 l1x 1.00 dead 0/8192
77
+ [epoch 55 step 1400] loss 432.3980 l_rec 221.5437 l_s 702.8479 l_aux 0.0000 l1x 1.00 dead 0/8192
78
+ [epoch 59 step 1500] loss 429.9404 l_rec 219.4620 l_s 701.5947 l_aux 0.0000 l1x 1.00 dead 0/8192
79
+ [epoch 63 step 1600] loss 428.2020 l_rec 218.0702 l_s 700.4395 l_aux 0.0000 l1x 1.00 dead 0/8192
80
+ [epoch 67 step 1700] loss 426.9943 l_rec 216.3623 l_s 702.1067 l_aux 0.0000 l1x 1.00 dead 0/8192
81
+ [epoch 71 step 1800] loss 426.0159 l_rec 215.0081 l_s 703.3591 l_aux 0.0000 l1x 1.00 dead 0/8192
82
+ [epoch 75 step 1900] loss 422.8452 l_rec 213.3365 l_s 698.3623 l_aux 0.0000 l1x 1.00 dead 0/8192
83
+ [epoch 79 step 2000] loss 422.1252 l_rec 211.9783 l_s 700.4897 l_aux 0.0000 l1x 1.00 dead 0/8192
84
+ [epoch 83 step 2100] loss 420.5292 l_rec 210.5843 l_s 699.8160 l_aux 0.0000 l1x 1.00 dead 0/8192
85
+ [epoch 87 step 2200] loss 418.2921 l_rec 209.5605 l_s 695.7720 l_aux 0.0000 l1x 1.00 dead 0/8192
86
+ [epoch 91 step 2300] loss 418.2229 l_rec 208.7839 l_s 698.1300 l_aux 0.0000 l1x 1.00 dead 0/8192
87
+ [epoch 95 step 2400] loss 417.1707 l_rec 208.2413 l_s 696.4312 l_aux 0.0000 l1x 1.00 dead 0/8192
88
+ [epoch 99 step 2500] loss 416.4127 l_rec 208.0586 l_s 694.5137 l_aux 0.0000 l1x 1.00 dead 0/8192
89
+ [step 2] EV=0.7289 L0=1150.1 (38.4s)
90
+ [step 2] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step2.pt
91
+ [step 4] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
92
+ [step 4] preprocess center_scale: mean-shift=0.003 scale=0.0228
93
+ [epoch 3 step 100] loss 760.4548 l_rec 760.3161 l_s 1.1671 l_aux 0.0000 l1x 0.40 dead 0/8192
94
+ [epoch 7 step 200] loss 695.5547 l_rec 690.8659 l_s 19.6349 l_aux 0.0000 l1x 0.80 dead 0/8192
95
+ [epoch 11 step 300] loss 468.2333 l_rec 269.4000 l_s 662.7775 l_aux 0.0000 l1x 1.00 dead 0/8192
96
+ [epoch 15 step 400] loss 458.0770 l_rec 254.2057 l_s 679.5710 l_aux 0.0000 l1x 1.00 dead 0/8192
97
+ [epoch 19 step 500] loss 454.0728 l_rec 246.9200 l_s 690.5092 l_aux 0.0000 l1x 1.00 dead 0/8192
98
+ [epoch 23 step 600] loss 450.4723 l_rec 242.1719 l_s 694.3348 l_aux 0.0000 l1x 1.00 dead 0/8192
99
+ [epoch 27 step 700] loss 447.8886 l_rec 238.7453 l_s 697.1443 l_aux 0.0000 l1x 1.00 dead 0/8192
100
+ [epoch 31 step 800] loss 445.7867 l_rec 235.5005 l_s 700.9538 l_aux 0.0000 l1x 1.00 dead 0/8192
101
+ [epoch 35 step 900] loss 443.7097 l_rec 232.9810 l_s 702.4290 l_aux 0.0000 l1x 1.00 dead 0/8192
102
+ [epoch 39 step 1000] loss 440.5602 l_rec 230.1463 l_s 701.3795 l_aux 0.0000 l1x 1.00 dead 0/8192
103
+ [epoch 43 step 1100] loss 438.2455 l_rec 227.8422 l_s 701.3444 l_aux 0.0000 l1x 1.00 dead 0/8192
104
+ [epoch 47 step 1200] loss 436.6012 l_rec 225.4586 l_s 703.8088 l_aux 0.0000 l1x 1.00 dead 0/8192
105
+ [epoch 51 step 1300] loss 433.7459 l_rec 223.1943 l_s 701.8389 l_aux 0.0000 l1x 1.00 dead 0/8192
106
+ [epoch 55 step 1400] loss 432.3973 l_rec 221.5453 l_s 702.8400 l_aux 0.0000 l1x 1.00 dead 0/8192
107
+ [epoch 59 step 1500] loss 429.9358 l_rec 219.4674 l_s 701.5614 l_aux 0.0000 l1x 1.00 dead 0/8192
108
+ [epoch 63 step 1600] loss 428.2131 l_rec 218.0721 l_s 700.4700 l_aux 0.0000 l1x 1.00 dead 0/8192
109
+ [epoch 67 step 1700] loss 427.0049 l_rec 216.3370 l_s 702.2263 l_aux 0.0000 l1x 1.00 dead 0/8192
110
+ [epoch 71 step 1800] loss 426.0134 l_rec 215.0079 l_s 703.3515 l_aux 0.0000 l1x 1.00 dead 0/8192
111
+ [epoch 75 step 1900] loss 422.8441 l_rec 213.3420 l_s 698.3401 l_aux 0.0000 l1x 1.00 dead 0/8192
112
+ [epoch 79 step 2000] loss 422.1180 l_rec 211.9921 l_s 700.4198 l_aux 0.0000 l1x 1.00 dead 0/8192
113
+ [epoch 83 step 2100] loss 420.5298 l_rec 210.5924 l_s 699.7910 l_aux 0.0000 l1x 1.00 dead 0/8192
114
+ [epoch 87 step 2200] loss 418.2770 l_rec 209.5766 l_s 695.6682 l_aux 0.0000 l1x 1.00 dead 0/8192
115
+ [epoch 91 step 2300] loss 418.2155 l_rec 208.7980 l_s 698.0585 l_aux 0.0000 l1x 1.00 dead 0/8192
116
+ [epoch 95 step 2400] loss 417.1678 l_rec 208.2535 l_s 696.3812 l_aux 0.0000 l1x 1.00 dead 0/8192
117
+ [epoch 99 step 2500] loss 416.4057 l_rec 208.0504 l_s 694.5175 l_aux 0.0000 l1x 1.00 dead 0/8192
118
+ [step 4] EV=0.7289 L0=1150.1 (38.5s)
119
+ [step 4] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step4.pt
120
+ [step 8] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
121
+ [step 8] preprocess center_scale: mean-shift=0.003 scale=0.0228
122
+ [epoch 3 step 100] loss 760.4558 l_rec 760.3172 l_s 1.1670 l_aux 0.0000 l1x 0.40 dead 0/8192
123
+ [epoch 7 step 200] loss 695.5470 l_rec 690.8573 l_s 19.6386 l_aux 0.0000 l1x 0.80 dead 0/8192
124
+ [epoch 11 step 300] loss 468.2439 l_rec 269.3944 l_s 662.8314 l_aux 0.0000 l1x 1.00 dead 0/8192
125
+ [epoch 15 step 400] loss 458.0811 l_rec 254.1956 l_s 679.6186 l_aux 0.0000 l1x 1.00 dead 0/8192
126
+ [epoch 19 step 500] loss 454.0691 l_rec 246.9263 l_s 690.4761 l_aux 0.0000 l1x 1.00 dead 0/8192
127
+ [epoch 23 step 600] loss 450.4771 l_rec 242.1594 l_s 694.3924 l_aux 0.0000 l1x 1.00 dead 0/8192
128
+ [epoch 27 step 700] loss 447.8939 l_rec 238.7539 l_s 697.1331 l_aux 0.0000 l1x 1.00 dead 0/8192
129
+ [epoch 31 step 800] loss 445.7833 l_rec 235.5070 l_s 700.9210 l_aux 0.0000 l1x 1.00 dead 0/8192
130
+ [epoch 35 step 900] loss 443.6945 l_rec 232.9973 l_s 702.3241 l_aux 0.0000 l1x 1.00 dead 0/8192
131
+ [epoch 39 step 1000] loss 440.5569 l_rec 230.1505 l_s 701.3549 l_aux 0.0000 l1x 1.00 dead 0/8192
132
+ [epoch 43 step 1100] loss 438.2603 l_rec 227.8279 l_s 701.4414 l_aux 0.0000 l1x 1.00 dead 0/8192
133
+ [epoch 47 step 1200] loss 436.6067 l_rec 225.4595 l_s 703.8241 l_aux 0.0000 l1x 1.00 dead 0/8192
134
+ [epoch 51 step 1300] loss 433.7474 l_rec 223.1858 l_s 701.8721 l_aux 0.0000 l1x 1.00 dead 0/8192
135
+ [epoch 55 step 1400] loss 432.4003 l_rec 221.5435 l_s 702.8558 l_aux 0.0000 l1x 1.00 dead 0/8192
136
+ [epoch 59 step 1500] loss 429.9407 l_rec 219.4559 l_s 701.6160 l_aux 0.0000 l1x 1.00 dead 0/8192
137
+ [epoch 63 step 1600] loss 428.2173 l_rec 218.0670 l_s 700.5007 l_aux 0.0000 l1x 1.00 dead 0/8192
138
+ [epoch 67 step 1700] loss 427.0011 l_rec 216.3527 l_s 702.1611 l_aux 0.0000 l1x 1.00 dead 0/8192
139
+ [epoch 71 step 1800] loss 426.0106 l_rec 215.0110 l_s 703.3322 l_aux 0.0000 l1x 1.00 dead 0/8192
140
+ [epoch 75 step 1900] loss 422.8394 l_rec 213.3398 l_s 698.3319 l_aux 0.0000 l1x 1.00 dead 0/8192
141
+ [epoch 79 step 2000] loss 422.1116 l_rec 212.0041 l_s 700.3583 l_aux 0.0000 l1x 1.00 dead 0/8192
142
+ [epoch 83 step 2100] loss 420.5238 l_rec 210.6019 l_s 699.7394 l_aux 0.0000 l1x 1.00 dead 0/8192
143
+ [epoch 87 step 2200] loss 418.3022 l_rec 209.5447 l_s 695.8582 l_aux 0.0000 l1x 1.00 dead 0/8192
144
+ [epoch 91 step 2300] loss 418.2136 l_rec 208.8034 l_s 698.0339 l_aux 0.0000 l1x 1.00 dead 0/8192
145
+ [epoch 95 step 2400] loss 417.1841 l_rec 208.2515 l_s 696.4419 l_aux 0.0000 l1x 1.00 dead 0/8192
146
+ [epoch 99 step 2500] loss 416.3970 l_rec 208.0618 l_s 694.4506 l_aux 0.0000 l1x 1.00 dead 0/8192
147
+ [step 8] EV=0.7289 L0=1150.1 (38.4s)
148
+ [step 8] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step8.pt
149
+ [step 16] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
150
+ [step 16] preprocess center_scale: mean-shift=0.003 scale=0.0228
151
+ [epoch 3 step 100] loss 760.4561 l_rec 760.3174 l_s 1.1673 l_aux 0.0000 l1x 0.40 dead 0/8192
152
+ [epoch 7 step 200] loss 695.5381 l_rec 690.8475 l_s 19.6424 l_aux 0.0000 l1x 0.80 dead 0/8192
153
+ [epoch 11 step 300] loss 468.2364 l_rec 269.3989 l_s 662.7916 l_aux 0.0000 l1x 1.00 dead 0/8192
154
+ [epoch 15 step 400] loss 458.0727 l_rec 254.2119 l_s 679.5360 l_aux 0.0000 l1x 1.00 dead 0/8192
155
+ [epoch 19 step 500] loss 454.0635 l_rec 246.9337 l_s 690.4326 l_aux 0.0000 l1x 1.00 dead 0/8192
156
+ [epoch 23 step 600] loss 450.4669 l_rec 242.1622 l_s 694.3489 l_aux 0.0000 l1x 1.00 dead 0/8192
157
+ [epoch 27 step 700] loss 447.8974 l_rec 238.7337 l_s 697.2121 l_aux 0.0000 l1x 1.00 dead 0/8192
158
+ [epoch 31 step 800] loss 445.7800 l_rec 235.5186 l_s 700.8713 l_aux 0.0000 l1x 1.00 dead 0/8192
159
+ [epoch 35 step 900] loss 443.7072 l_rec 232.9761 l_s 702.4366 l_aux 0.0000 l1x 1.00 dead 0/8192
160
+ [epoch 39 step 1000] loss 440.5675 l_rec 230.1243 l_s 701.4772 l_aux 0.0000 l1x 1.00 dead 0/8192
161
+ [epoch 43 step 1100] loss 438.2390 l_rec 227.8392 l_s 701.3325 l_aux 0.0000 l1x 1.00 dead 0/8192
162
+ [epoch 47 step 1200] loss 436.6060 l_rec 225.4460 l_s 703.8668 l_aux 0.0000 l1x 1.00 dead 0/8192
163
+ [epoch 51 step 1300] loss 433.7491 l_rec 223.1839 l_s 701.8842 l_aux 0.0000 l1x 1.00 dead 0/8192
164
+ [epoch 55 step 1400] loss 432.4088 l_rec 221.5304 l_s 702.9281 l_aux 0.0000 l1x 1.00 dead 0/8192
165
+ [epoch 59 step 1500] loss 429.9540 l_rec 219.4434 l_s 701.7017 l_aux 0.0000 l1x 1.00 dead 0/8192
166
+ [epoch 63 step 1600] loss 428.1972 l_rec 218.0775 l_s 700.3988 l_aux 0.0000 l1x 1.00 dead 0/8192
167
+ [epoch 67 step 1700] loss 426.9910 l_rec 216.3561 l_s 702.1163 l_aux 0.0000 l1x 1.00 dead 0/8192
168
+ [epoch 71 step 1800] loss 426.0088 l_rec 215.0067 l_s 703.3405 l_aux 0.0000 l1x 1.00 dead 0/8192
169
+ [epoch 75 step 1900] loss 422.8248 l_rec 213.3554 l_s 698.2311 l_aux 0.0000 l1x 1.00 dead 0/8192
170
+ [epoch 79 step 2000] loss 422.1181 l_rec 212.0002 l_s 700.3929 l_aux 0.0000 l1x 1.00 dead 0/8192
171
+ [epoch 83 step 2100] loss 420.5402 l_rec 210.5772 l_s 699.8766 l_aux 0.0000 l1x 1.00 dead 0/8192
172
+ [epoch 87 step 2200] loss 418.3077 l_rec 209.5741 l_s 695.7787 l_aux 0.0000 l1x 1.00 dead 0/8192
173
+ [epoch 91 step 2300] loss 418.2187 l_rec 208.7874 l_s 698.1041 l_aux 0.0000 l1x 1.00 dead 0/8192
174
+ [epoch 95 step 2400] loss 417.1819 l_rec 208.2390 l_s 696.4765 l_aux 0.0000 l1x 1.00 dead 0/8192
175
+ [epoch 99 step 2500] loss 416.4051 l_rec 208.0755 l_s 694.4318 l_aux 0.0000 l1x 1.00 dead 0/8192
176
+ [step 16] EV=0.7289 L0=1150.1 (38.9s)
177
+ [step 16] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step16.pt
178
+ [step 32] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
179
+ [step 32] preprocess center_scale: mean-shift=0.004 scale=0.0228
180
+ [epoch 3 step 100] loss 760.4551 l_rec 760.3162 l_s 1.1687 l_aux 0.0000 l1x 0.40 dead 0/8192
181
+ [epoch 7 step 200] loss 695.5250 l_rec 690.8338 l_s 19.6449 l_aux 0.0000 l1x 0.80 dead 0/8192
182
+ [epoch 11 step 300] loss 468.2241 l_rec 269.4258 l_s 662.6608 l_aux 0.0000 l1x 1.00 dead 0/8192
183
+ [epoch 15 step 400] loss 458.0741 l_rec 254.2050 l_s 679.5634 l_aux 0.0000 l1x 1.00 dead 0/8192
184
+ [epoch 19 step 500] loss 454.0776 l_rec 246.9102 l_s 690.5581 l_aux 0.0000 l1x 1.00 dead 0/8192
185
+ [epoch 23 step 600] loss 450.4738 l_rec 242.1669 l_s 694.3562 l_aux 0.0000 l1x 1.00 dead 0/8192
186
+ [epoch 27 step 700] loss 447.8953 l_rec 238.7349 l_s 697.2014 l_aux 0.0000 l1x 1.00 dead 0/8192
187
+ [epoch 31 step 800] loss 445.7839 l_rec 235.5098 l_s 700.9138 l_aux 0.0000 l1x 1.00 dead 0/8192
188
+ [epoch 35 step 900] loss 443.7005 l_rec 232.9933 l_s 702.3574 l_aux 0.0000 l1x 1.00 dead 0/8192
189
+ [epoch 39 step 1000] loss 440.5628 l_rec 230.1317 l_s 701.4371 l_aux 0.0000 l1x 1.00 dead 0/8192
190
+ [epoch 43 step 1100] loss 438.2335 l_rec 227.8497 l_s 701.2793 l_aux 0.0000 l1x 1.00 dead 0/8192
191
+ [epoch 47 step 1200] loss 436.6121 l_rec 225.4347 l_s 703.9247 l_aux 0.0000 l1x 1.00 dead 0/8192
192
+ [epoch 51 step 1300] loss 433.7444 l_rec 223.1962 l_s 701.8276 l_aux 0.0000 l1x 1.00 dead 0/8192
193
+ [epoch 55 step 1400] loss 432.3934 l_rec 221.5427 l_s 702.8357 l_aux 0.0000 l1x 1.00 dead 0/8192
194
+ [epoch 59 step 1500] loss 429.9392 l_rec 219.4641 l_s 701.5835 l_aux 0.0000 l1x 1.00 dead 0/8192
195
+ [epoch 63 step 1600] loss 428.2032 l_rec 218.0660 l_s 700.4573 l_aux 0.0000 l1x 1.00 dead 0/8192
196
+ [epoch 67 step 1700] loss 426.9786 l_rec 216.3650 l_s 702.0454 l_aux 0.0000 l1x 1.00 dead 0/8192
197
+ [epoch 71 step 1800] loss 426.0150 l_rec 214.9917 l_s 703.4106 l_aux 0.0000 l1x 1.00 dead 0/8192
198
+ [epoch 75 step 1900] loss 422.7785 l_rec 213.3622 l_s 698.0542 l_aux 0.0000 l1x 1.00 dead 0/8192
199
+ [epoch 79 step 2000] loss 422.1183 l_rec 211.9862 l_s 700.4403 l_aux 0.0000 l1x 1.00 dead 0/8192
200
+ [epoch 83 step 2100] loss 420.5203 l_rec 210.5933 l_s 699.7565 l_aux 0.0000 l1x 1.00 dead 0/8192
201
+ [epoch 87 step 2200] loss 418.3167 l_rec 209.5382 l_s 695.9282 l_aux 0.0000 l1x 1.00 dead 0/8192
202
+ [epoch 91 step 2300] loss 418.1925 l_rec 208.7971 l_s 697.9846 l_aux 0.0000 l1x 1.00 dead 0/8192
203
+ [epoch 95 step 2400] loss 417.1999 l_rec 208.2266 l_s 696.5776 l_aux 0.0000 l1x 1.00 dead 0/8192
204
+ [epoch 99 step 2500] loss 416.3907 l_rec 208.0703 l_s 694.4012 l_aux 0.0000 l1x 1.00 dead 0/8192
205
+ [step 32] EV=0.7289 L0=1150.1 (38.6s)
206
+ [step 32] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step32.pt
207
+ [step 64] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
208
+ [step 64] preprocess center_scale: mean-shift=0.015 scale=0.0228
209
+ [epoch 3 step 100] loss 760.4854 l_rec 760.3465 l_s 1.1686 l_aux 0.0000 l1x 0.40 dead 0/8192
210
+ [epoch 7 step 200] loss 695.5303 l_rec 690.8440 l_s 19.6244 l_aux 0.0000 l1x 0.80 dead 0/8192
211
+ [epoch 11 step 300] loss 468.2451 l_rec 269.4274 l_s 662.7256 l_aux 0.0000 l1x 1.00 dead 0/8192
212
+ [epoch 15 step 400] loss 458.1093 l_rec 254.2168 l_s 679.6416 l_aux 0.0000 l1x 1.00 dead 0/8192
213
+ [epoch 19 step 500] loss 454.0862 l_rec 246.8946 l_s 690.6385 l_aux 0.0000 l1x 1.00 dead 0/8192
214
+ [epoch 23 step 600] loss 450.4788 l_rec 242.1711 l_s 694.3591 l_aux 0.0000 l1x 1.00 dead 0/8192
215
+ [epoch 27 step 700] loss 447.8814 l_rec 238.7346 l_s 697.1558 l_aux 0.0000 l1x 1.00 dead 0/8192
216
+ [epoch 31 step 800] loss 445.8001 l_rec 235.5198 l_s 700.9343 l_aux 0.0000 l1x 1.00 dead 0/8192
217
+ [epoch 35 step 900] loss 443.7321 l_rec 232.9727 l_s 702.5314 l_aux 0.0000 l1x 1.00 dead 0/8192
218
+ [epoch 39 step 1000] loss 440.5579 l_rec 230.1348 l_s 701.4102 l_aux 0.0000 l1x 1.00 dead 0/8192
219
+ [epoch 43 step 1100] loss 438.2473 l_rec 227.8575 l_s 701.2992 l_aux 0.0000 l1x 1.00 dead 0/8192
220
+ [epoch 47 step 1200] loss 436.6013 l_rec 225.4315 l_s 703.8994 l_aux 0.0000 l1x 1.00 dead 0/8192
221
+ [epoch 51 step 1300] loss 433.7101 l_rec 223.1986 l_s 701.7051 l_aux 0.0000 l1x 1.00 dead 0/8192
222
+ [epoch 55 step 1400] loss 432.4117 l_rec 221.5196 l_s 702.9737 l_aux 0.0000 l1x 1.00 dead 0/8192
223
+ [epoch 59 step 1500] loss 429.9248 l_rec 219.4503 l_s 701.5817 l_aux 0.0000 l1x 1.00 dead 0/8192
224
+ [epoch 63 step 1600] loss 428.2202 l_rec 218.0621 l_s 700.5272 l_aux 0.0000 l1x 1.00 dead 0/8192
225
+ [epoch 67 step 1700] loss 426.9976 l_rec 216.3706 l_s 702.0901 l_aux 0.0000 l1x 1.00 dead 0/8192
226
+ [epoch 71 step 1800] loss 425.9592 l_rec 215.0018 l_s 703.1913 l_aux 0.0000 l1x 1.00 dead 0/8192
227
+ [epoch 75 step 1900] loss 422.8094 l_rec 213.3825 l_s 698.0897 l_aux 0.0000 l1x 1.00 dead 0/8192
228
+ [epoch 79 step 2000] loss 422.1139 l_rec 212.0237 l_s 700.3006 l_aux 0.0000 l1x 1.00 dead 0/8192
229
+ [epoch 83 step 2100] loss 420.4716 l_rec 210.6061 l_s 699.5519 l_aux 0.0000 l1x 1.00 dead 0/8192
230
+ [epoch 87 step 2200] loss 418.3567 l_rec 209.5072 l_s 696.1650 l_aux 0.0000 l1x 1.00 dead 0/8192
231
+ [epoch 91 step 2300] loss 418.1878 l_rec 208.8147 l_s 697.9102 l_aux 0.0000 l1x 1.00 dead 0/8192
232
+ [epoch 95 step 2400] loss 417.2208 l_rec 208.2696 l_s 696.5039 l_aux 0.0000 l1x 1.00 dead 0/8192
233
+ [epoch 99 step 2500] loss 416.3930 l_rec 208.0202 l_s 694.5759 l_aux 0.0000 l1x 1.00 dead 0/8192
234
+ [step 64] EV=0.7289 L0=1150.1 (38.5s)
235
+ [step 64] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step64.pt
236
+ [step 128] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
237
+ [step 128] preprocess center_scale: mean-shift=0.059 scale=0.0228
238
+ [epoch 3 step 100] loss 760.4408 l_rec 760.3014 l_s 1.1733 l_aux 0.0000 l1x 0.40 dead 0/8192
239
+ [epoch 7 step 200] loss 694.8250 l_rec 690.0862 l_s 19.8444 l_aux 0.0000 l1x 0.80 dead 0/8192
240
+ [epoch 11 step 300] loss 468.3123 l_rec 270.2250 l_s 660.2908 l_aux 0.0000 l1x 1.00 dead 0/8192
241
+ [epoch 15 step 400] loss 458.0602 l_rec 254.2044 l_s 679.5194 l_aux 0.0000 l1x 1.00 dead 0/8192
242
+ [epoch 19 step 500] loss 453.9016 l_rec 246.8609 l_s 690.1357 l_aux 0.0000 l1x 1.00 dead 0/8192
243
+ [epoch 23 step 600] loss 450.2691 l_rec 242.0764 l_s 693.9758 l_aux 0.0000 l1x 1.00 dead 0/8192
244
+ [epoch 27 step 700] loss 447.7215 l_rec 238.6661 l_s 696.8513 l_aux 0.0000 l1x 1.00 dead 0/8192
245
+ [epoch 31 step 800] loss 445.6384 l_rec 235.3931 l_s 700.8177 l_aux 0.0000 l1x 1.00 dead 0/8192
246
+ [epoch 35 step 900] loss 443.5186 l_rec 232.8480 l_s 702.2354 l_aux 0.0000 l1x 1.00 dead 0/8192
247
+ [epoch 39 step 1000] loss 440.2974 l_rec 229.9970 l_s 701.0010 l_aux 0.0000 l1x 1.00 dead 0/8192
248
+ [epoch 43 step 1100] loss 437.9800 l_rec 227.8564 l_s 700.4120 l_aux 0.0000 l1x 1.00 dead 0/8192
249
+ [epoch 47 step 1200] loss 436.4210 l_rec 225.3456 l_s 703.5844 l_aux 0.0000 l1x 1.00 dead 0/8192
250
+ [epoch 51 step 1300] loss 433.3655 l_rec 223.1081 l_s 700.8580 l_aux 0.0000 l1x 1.00 dead 0/8192
251
+ [epoch 55 step 1400] loss 432.3748 l_rec 221.4286 l_s 703.1541 l_aux 0.0000 l1x 1.00 dead 0/8192
252
+ [epoch 59 step 1500] loss 429.4362 l_rec 219.3235 l_s 700.3758 l_aux 0.0000 l1x 1.00 dead 0/8192
253
+ [epoch 63 step 1600] loss 427.9319 l_rec 217.9848 l_s 699.8235 l_aux 0.0000 l1x 1.00 dead 0/8192
254
+ [epoch 67 step 1700] loss 426.7894 l_rec 216.2594 l_s 701.7665 l_aux 0.0000 l1x 1.00 dead 0/8192
255
+ [epoch 71 step 1800] loss 425.7352 l_rec 214.8456 l_s 702.9653 l_aux 0.0000 l1x 1.00 dead 0/8192
256
+ [epoch 75 step 1900] loss 422.6576 l_rec 213.1798 l_s 698.2593 l_aux 0.0000 l1x 1.00 dead 0/8192
257
+ [epoch 79 step 2000] loss 422.0034 l_rec 211.8541 l_s 700.4976 l_aux 0.0000 l1x 1.00 dead 0/8192
258
+ [epoch 83 step 2100] loss 420.0692 l_rec 210.4810 l_s 698.6272 l_aux 0.0000 l1x 1.00 dead 0/8192
259
+ [epoch 87 step 2200] loss 418.0375 l_rec 209.4419 l_s 695.3186 l_aux 0.0000 l1x 1.00 dead 0/8192
260
+ [epoch 91 step 2300] loss 418.0702 l_rec 208.6158 l_s 698.1811 l_aux 0.0000 l1x 1.00 dead 0/8192
261
+ [epoch 95 step 2400] loss 417.0173 l_rec 208.0962 l_s 696.4037 l_aux 0.0000 l1x 1.00 dead 0/8192
262
+ [epoch 99 step 2500] loss 416.1335 l_rec 207.8891 l_s 694.1479 l_aux 0.0000 l1x 1.00 dead 0/8192
263
+ [step 128] EV=0.7290 L0=1149.6 (38.4s)
264
+ [step 128] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step128.pt
265
+ [step 256] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
266
+ [step 256] preprocess center_scale: mean-shift=0.131 scale=0.0231
267
+ [epoch 3 step 100] loss 760.1663 l_rec 760.0106 l_s 1.3107 l_aux 0.0000 l1x 0.40 dead 0/8192
268
+ [epoch 7 step 200] loss 666.5573 l_rec 656.4554 l_s 42.3024 l_aux 0.0000 l1x 0.80 dead 0/8192
269
+ [epoch 11 step 300] loss 473.8857 l_rec 301.3712 l_s 575.0483 l_aux 0.0000 l1x 1.00 dead 0/8192
270
+ [epoch 15 step 400] loss 455.1253 l_rec 255.7930 l_s 664.4409 l_aux 0.0000 l1x 1.00 dead 0/8192
271
+ [epoch 19 step 500] loss 450.7234 l_rec 246.3354 l_s 681.2932 l_aux 0.0000 l1x 1.00 dead 0/8192
272
+ [epoch 23 step 600] loss 446.3193 l_rec 241.0368 l_s 684.2748 l_aux 0.0000 l1x 1.00 dead 0/8192
273
+ [epoch 27 step 700] loss 444.3858 l_rec 237.3688 l_s 690.0568 l_aux 0.0000 l1x 1.00 dead 0/8192
274
+ [epoch 31 step 800] loss 441.7399 l_rec 234.0233 l_s 692.3887 l_aux 0.0000 l1x 1.00 dead 0/8192
275
+ [epoch 35 step 900] loss 439.7041 l_rec 231.3902 l_s 694.3796 l_aux 0.0000 l1x 1.00 dead 0/8192
276
+ [epoch 39 step 1000] loss 436.5400 l_rec 228.5362 l_s 693.3457 l_aux 0.0000 l1x 1.00 dead 0/8192
277
+ [epoch 43 step 1100] loss 434.1518 l_rec 226.0798 l_s 693.5732 l_aux 0.0000 l1x 1.00 dead 0/8192
278
+ [epoch 47 step 1200] loss 432.3850 l_rec 223.5822 l_s 696.0092 l_aux 0.0000 l1x 1.00 dead 0/8192
279
+ [epoch 51 step 1300] loss 429.1430 l_rec 221.2554 l_s 692.9585 l_aux 0.0000 l1x 1.00 dead 0/8192
280
+ [epoch 55 step 1400] loss 428.7843 l_rec 219.5893 l_s 697.3165 l_aux 0.0000 l1x 1.00 dead 0/8192
281
+ [epoch 59 step 1500] loss 424.8134 l_rec 217.4993 l_s 691.0471 l_aux 0.0000 l1x 1.00 dead 0/8192
282
+ [epoch 63 step 1600] loss 423.2327 l_rec 215.9136 l_s 691.0638 l_aux 0.0000 l1x 1.00 dead 0/8192
283
+ [epoch 67 step 1700] loss 422.6964 l_rec 214.4421 l_s 694.1807 l_aux 0.0000 l1x 1.00 dead 0/8192
284
+ [epoch 71 step 1800] loss 421.7040 l_rec 213.0388 l_s 695.5505 l_aux 0.0000 l1x 1.00 dead 0/8192
285
+ [epoch 75 step 1900] loss 418.3461 l_rec 211.0828 l_s 690.8777 l_aux 0.0000 l1x 1.00 dead 0/8192
286
+ [epoch 79 step 2000] loss 418.0120 l_rec 209.7280 l_s 694.2800 l_aux 0.0000 l1x 1.00 dead 0/8192
287
+ [epoch 83 step 2100] loss 415.3140 l_rec 208.4247 l_s 689.6309 l_aux 0.0000 l1x 1.00 dead 0/8192
288
+ [epoch 87 step 2200] loss 413.6560 l_rec 207.3763 l_s 687.5989 l_aux 0.0000 l1x 1.00 dead 0/8192
289
+ [epoch 91 step 2300] loss 414.1241 l_rec 206.5059 l_s 692.0607 l_aux 0.0000 l1x 1.00 dead 0/8192
290
+ [epoch 95 step 2400] loss 412.6747 l_rec 205.9826 l_s 688.9735 l_aux 0.0000 l1x 1.00 dead 0/8192
291
+ [epoch 99 step 2500] loss 411.8680 l_rec 205.7821 l_s 686.9528 l_aux 0.0000 l1x 1.00 dead 0/8192
292
+ [step 256] EV=0.7317 L0=1142.4 (38.6s)
293
+ [step 256] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step256.pt
294
+ [step 512] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
295
+ [step 512] preprocess center_scale: mean-shift=0.213 scale=0.0257
296
+ [epoch 3 step 100] loss 747.9344 l_rec 747.5690 l_s 3.0758 l_aux 0.0000 l1x 0.40 dead 0/8192
297
+ [epoch 7 step 200] loss 545.6375 l_rec 490.7026 l_s 230.0454 l_aux 0.0000 l1x 0.80 dead 0/8192
298
+ [epoch 11 step 300] loss 478.7583 l_rec 359.9893 l_s 395.8965 l_aux 0.0000 l1x 1.00 dead 0/8192
299
+ [epoch 15 step 400] loss 437.5027 l_rec 289.3768 l_s 493.7529 l_aux 0.0000 l1x 1.00 dead 0/8192
300
+ [epoch 19 step 500] loss 417.6176 l_rec 248.9130 l_s 562.3485 l_aux 0.0000 l1x 1.00 dead 0/8192
301
+ [epoch 23 step 600] loss 414.0464 l_rec 232.7734 l_s 604.2433 l_aux 0.0000 l1x 1.00 dead 0/8192
302
+ [epoch 27 step 700] loss 412.9706 l_rec 225.4388 l_s 625.1058 l_aux 0.0000 l1x 1.00 dead 0/8192
303
+ [epoch 31 step 800] loss 407.1723 l_rec 219.7163 l_s 624.8534 l_aux 0.0000 l1x 1.00 dead 0/8192
304
+ [epoch 35 step 900] loss 405.6177 l_rec 216.2735 l_s 631.1475 l_aux 0.0000 l1x 1.00 dead 0/8192
305
+ [epoch 39 step 1000] loss 401.8006 l_rec 212.6550 l_s 630.4851 l_aux 0.0000 l1x 1.00 dead 0/8192
306
+ [epoch 43 step 1100] loss 398.0692 l_rec 209.7099 l_s 627.8644 l_aux 0.0000 l1x 1.00 dead 0/8192
307
+ [epoch 47 step 1200] loss 398.1291 l_rec 207.3442 l_s 635.9496 l_aux 0.0000 l1x 1.00 dead 0/8192
308
+ [epoch 51 step 1300] loss 393.8828 l_rec 204.9331 l_s 629.8323 l_aux 0.0000 l1x 1.00 dead 0/8192
309
+ [epoch 55 step 1400] loss 393.0395 l_rec 202.7346 l_s 634.3497 l_aux 0.0000 l1x 1.00 dead 0/8192
310
+ [epoch 59 step 1500] loss 389.0173 l_rec 200.3399 l_s 628.9248 l_aux 0.0000 l1x 1.00 dead 0/8192
311
+ [epoch 63 step 1600] loss 385.8784 l_rec 198.5999 l_s 624.2614 l_aux 0.0000 l1x 1.00 dead 0/8192
312
+ [epoch 67 step 1700] loss 387.0005 l_rec 197.2491 l_s 632.5050 l_aux 0.0000 l1x 1.00 dead 0/8192
313
+ [epoch 71 step 1800] loss 384.8597 l_rec 195.5052 l_s 631.1816 l_aux 0.0000 l1x 1.00 dead 0/8192
314
+ [epoch 75 step 1900] loss 381.6938 l_rec 193.7720 l_s 626.4059 l_aux 0.0000 l1x 1.00 dead 0/8192
315
+ [epoch 79 step 2000] loss 382.1370 l_rec 192.5943 l_s 631.8089 l_aux 0.0000 l1x 1.00 dead 0/8192
316
+ [epoch 83 step 2100] loss 380.1515 l_rec 191.2267 l_s 629.7493 l_aux 0.0000 l1x 1.00 dead 0/8192
317
+ [epoch 87 step 2200] loss 376.8493 l_rec 189.9150 l_s 623.1144 l_aux 0.0000 l1x 1.00 dead 0/8192
318
+ [epoch 91 step 2300] loss 377.8456 l_rec 189.4662 l_s 627.9313 l_aux 0.0000 l1x 1.00 dead 0/8192
319
+ [epoch 95 step 2400] loss 376.1934 l_rec 188.6144 l_s 625.2634 l_aux 0.0000 l1x 1.00 dead 0/8192
320
+ [epoch 99 step 2500] loss 374.6293 l_rec 188.6015 l_s 620.0929 l_aux 0.0000 l1x 1.00 dead 0/8192
321
+ [step 512] EV=0.7543 L0=1086.6 (38.4s)
322
+ [step 512] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step512.pt
323
+ [step 1000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
324
+ [step 1000] preprocess center_scale: mean-shift=0.276 scale=0.0325
325
+ [epoch 3 step 100] loss 732.6374 l_rec 732.0483 l_s 4.9586 l_aux 0.0000 l1x 0.40 dead 0/8192
326
+ [epoch 7 step 200] loss 480.6347 l_rec 409.5700 l_s 297.5906 l_aux 0.0000 l1x 0.80 dead 0/8192
327
+ [epoch 11 step 300] loss 419.4728 l_rec 313.4902 l_s 353.2754 l_aux 0.0000 l1x 1.00 dead 0/8192
328
+ [epoch 15 step 400] loss 388.1238 l_rec 270.4863 l_s 392.1248 l_aux 0.0000 l1x 1.00 dead 0/8192
329
+ [epoch 19 step 500] loss 371.3777 l_rec 241.6649 l_s 432.3760 l_aux 0.0000 l1x 1.00 dead 0/8192
330
+ [epoch 23 step 600] loss 366.6693 l_rec 224.6686 l_s 473.3357 l_aux 0.0000 l1x 1.00 dead 0/8192
331
+ [epoch 27 step 700] loss 366.9374 l_rec 213.2239 l_s 512.3782 l_aux 0.0000 l1x 1.00 dead 0/8192
332
+ [epoch 31 step 800] loss 359.5863 l_rec 203.6777 l_s 519.6954 l_aux 0.0000 l1x 1.00 dead 0/8192
333
+ [epoch 35 step 900] loss 356.7603 l_rec 197.2643 l_s 531.6532 l_aux 0.0000 l1x 1.00 dead 0/8192
334
+ [epoch 39 step 1000] loss 355.3755 l_rec 192.5407 l_s 542.7830 l_aux 0.0000 l1x 1.00 dead 0/8192
335
+ [epoch 43 step 1100] loss 348.2111 l_rec 187.2364 l_s 536.5820 l_aux 0.0000 l1x 1.00 dead 0/8192
336
+ [epoch 47 step 1200] loss 347.7171 l_rec 185.1862 l_s 541.7698 l_aux 0.0000 l1x 1.00 dead 0/8192
337
+ [epoch 51 step 1300] loss 344.7083 l_rec 181.9315 l_s 542.5892 l_aux 0.0000 l1x 1.00 dead 0/8192
338
+ [epoch 55 step 1400] loss 345.2197 l_rec 179.4756 l_s 552.4800 l_aux 0.0000 l1x 1.00 dead 0/8192
339
+ [epoch 59 step 1500] loss 340.5815 l_rec 176.7815 l_s 545.9999 l_aux 0.0000 l1x 1.00 dead 0/8192
340
+ [epoch 63 step 1600] loss 338.4995 l_rec 175.1846 l_s 544.3828 l_aux 0.0000 l1x 1.00 dead 0/8192
341
+ [epoch 67 step 1700] loss 340.1748 l_rec 173.3462 l_s 556.0953 l_aux 0.0000 l1x 1.00 dead 0/8192
342
+ [epoch 71 step 1800] loss 335.4354 l_rec 171.1983 l_s 547.4568 l_aux 0.0000 l1x 1.00 dead 0/8192
343
+ [epoch 75 step 1900] loss 333.0423 l_rec 169.4483 l_s 545.3132 l_aux 0.0000 l1x 1.00 dead 0/8192
344
+ [epoch 79 step 2000] loss 333.0328 l_rec 168.3450 l_s 548.9596 l_aux 0.0000 l1x 1.00 dead 0/8192
345
+ [epoch 83 step 2100] loss 332.5890 l_rec 167.0401 l_s 551.8297 l_aux 0.0000 l1x 1.00 dead 0/8192
346
+ [epoch 87 step 2200] loss 326.5447 l_rec 165.1005 l_s 538.1473 l_aux 0.0000 l1x 1.00 dead 0/8192
347
+ [epoch 91 step 2300] loss 327.8544 l_rec 165.1454 l_s 542.3635 l_aux 0.0000 l1x 1.00 dead 0/8192
348
+ [epoch 95 step 2400] loss 324.4103 l_rec 163.9068 l_s 535.0116 l_aux 0.0000 l1x 1.00 dead 0/8192
349
+ [epoch 99 step 2500] loss 323.7086 l_rec 163.6606 l_s 533.4937 l_aux 0.0000 l1x 1.00 dead 0/8192
350
+ [step 1000] EV=0.7862 L0=996.9 (38.8s)
351
+ [step 1000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step1000.pt
352
+ [step 2000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
353
+ [step 2000] preprocess center_scale: mean-shift=0.245 scale=0.0394
354
+ [epoch 3 step 100] loss 731.7496 l_rec 731.1732 l_s 4.8518 l_aux 0.0000 l1x 0.40 dead 0/8192
355
+ [epoch 7 step 200] loss 497.2430 l_rec 424.9264 l_s 302.8332 l_aux 0.0000 l1x 0.80 dead 0/8192
356
+ [epoch 11 step 300] loss 412.5962 l_rec 296.1410 l_s 388.1839 l_aux 0.0000 l1x 1.00 dead 0/8192
357
+ [epoch 15 step 400] loss 376.2416 l_rec 249.6217 l_s 422.0665 l_aux 0.0000 l1x 1.00 dead 0/8192
358
+ [epoch 19 step 500] loss 359.9390 l_rec 223.6637 l_s 454.2508 l_aux 0.0000 l1x 1.00 dead 0/8192
359
+ [epoch 23 step 600] loss 354.7119 l_rec 209.9524 l_s 482.5319 l_aux 0.0000 l1x 1.00 dead 0/8192
360
+ [epoch 27 step 700] loss 355.6157 l_rec 201.6862 l_s 513.0980 l_aux 0.0000 l1x 1.00 dead 0/8192
361
+ [epoch 31 step 800] loss 349.8182 l_rec 194.3371 l_s 518.2701 l_aux 0.0000 l1x 1.00 dead 0/8192
362
+ [epoch 35 step 900] loss 342.3737 l_rec 186.9730 l_s 518.0023 l_aux 0.0000 l1x 1.00 dead 0/8192
363
+ [epoch 39 step 1000] loss 346.1694 l_rec 184.7088 l_s 538.2017 l_aux 0.0000 l1x 1.00 dead 0/8192
364
+ [epoch 43 step 1100] loss 334.6384 l_rec 178.0774 l_s 521.8702 l_aux 0.0000 l1x 1.00 dead 0/8192
365
+ [epoch 47 step 1200] loss 333.2682 l_rec 175.9952 l_s 524.2435 l_aux 0.0000 l1x 1.00 dead 0/8192
366
+ [epoch 51 step 1300] loss 331.6114 l_rec 172.6459 l_s 529.8848 l_aux 0.0000 l1x 1.00 dead 0/8192
367
+ [epoch 55 step 1400] loss 333.9041 l_rec 171.0789 l_s 542.7505 l_aux 0.0000 l1x 1.00 dead 0/8192
368
+ [epoch 59 step 1500] loss 327.9637 l_rec 167.4702 l_s 534.9784 l_aux 0.0000 l1x 1.00 dead 0/8192
369
+ [epoch 63 step 1600] loss 326.3385 l_rec 165.7961 l_s 535.1414 l_aux 0.0000 l1x 1.00 dead 0/8192
370
+ [epoch 67 step 1700] loss 327.1721 l_rec 164.5744 l_s 541.9923 l_aux 0.0000 l1x 1.00 dead 0/8192
371
+ [epoch 71 step 1800] loss 323.6122 l_rec 162.0108 l_s 538.6711 l_aux 0.0000 l1x 1.00 dead 0/8192
372
+ [epoch 75 step 1900] loss 319.7923 l_rec 159.7250 l_s 533.5578 l_aux 0.0000 l1x 1.00 dead 0/8192
373
+ [epoch 79 step 2000] loss 320.3025 l_rec 159.1009 l_s 537.3387 l_aux 0.0000 l1x 1.00 dead 0/8192
374
+ [epoch 83 step 2100] loss 319.8980 l_rec 157.6922 l_s 540.6862 l_aux 0.0000 l1x 1.00 dead 0/8192
375
+ [epoch 87 step 2200] loss 312.6612 l_rec 154.7510 l_s 526.3672 l_aux 0.0000 l1x 1.00 dead 0/8192
376
+ [epoch 91 step 2300] loss 315.1534 l_rec 155.3063 l_s 532.8236 l_aux 0.0000 l1x 1.00 dead 0/8192
377
+ [epoch 95 step 2400] loss 309.1252 l_rec 153.3327 l_s 519.3083 l_aux 0.0000 l1x 1.00 dead 0/8192
378
+ [epoch 99 step 2500] loss 310.0103 l_rec 153.3199 l_s 522.3014 l_aux 0.0000 l1x 1.00 dead 0/8192
379
+ [step 2000] EV=0.7993 L0=969.2 (38.4s)
380
+ [step 2000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step2000.pt
381
+ [step 3000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
382
+ [step 3000] preprocess center_scale: mean-shift=0.230 scale=0.0420
383
+ [epoch 3 step 100] loss 737.0491 l_rec 736.5306 l_s 4.3643 l_aux 0.0000 l1x 0.40 dead 0/8192
384
+ [epoch 7 step 200] loss 514.4067 l_rec 446.4767 l_s 284.4640 l_aux 0.0000 l1x 0.80 dead 0/8192
385
+ [epoch 11 step 300] loss 422.6810 l_rec 303.9492 l_s 395.7729 l_aux 0.0000 l1x 1.00 dead 0/8192
386
+ [epoch 15 step 400] loss 384.0181 l_rec 251.4632 l_s 441.8497 l_aux 0.0000 l1x 1.00 dead 0/8192
387
+ [epoch 19 step 500] loss 367.2769 l_rec 223.7237 l_s 478.5106 l_aux 0.0000 l1x 1.00 dead 0/8192
388
+ [epoch 23 step 600] loss 361.9778 l_rec 209.5129 l_s 508.2165 l_aux 0.0000 l1x 1.00 dead 0/8192
389
+ [epoch 27 step 700] loss 361.7198 l_rec 201.1429 l_s 535.2565 l_aux 0.0000 l1x 1.00 dead 0/8192
390
+ [epoch 31 step 800] loss 357.1411 l_rec 194.6059 l_s 541.7839 l_aux 0.0000 l1x 1.00 dead 0/8192
391
+ [epoch 35 step 900] loss 346.6657 l_rec 186.2998 l_s 534.5530 l_aux 0.0000 l1x 1.00 dead 0/8192
392
+ [epoch 39 step 1000] loss 353.2422 l_rec 185.1582 l_s 560.2803 l_aux 0.0000 l1x 1.00 dead 0/8192
393
+ [epoch 43 step 1100] loss 340.1267 l_rec 178.2385 l_s 539.6272 l_aux 0.0000 l1x 1.00 dead 0/8192
394
+ [epoch 47 step 1200] loss 338.1505 l_rec 175.7970 l_s 541.1784 l_aux 0.0000 l1x 1.00 dead 0/8192
395
+ [epoch 51 step 1300] loss 336.6627 l_rec 172.8397 l_s 546.0767 l_aux 0.0000 l1x 1.00 dead 0/8192
396
+ [epoch 55 step 1400] loss 339.1928 l_rec 171.3967 l_s 559.3204 l_aux 0.0000 l1x 1.00 dead 0/8192
397
+ [epoch 59 step 1500] loss 333.9812 l_rec 167.9312 l_s 553.4998 l_aux 0.0000 l1x 1.00 dead 0/8192
398
+ [epoch 63 step 1600] loss 331.3563 l_rec 165.9082 l_s 551.4937 l_aux 0.0000 l1x 1.00 dead 0/8192
399
+ [epoch 67 step 1700] loss 331.4246 l_rec 164.3187 l_s 557.0197 l_aux 0.0000 l1x 1.00 dead 0/8192
400
+ [epoch 71 step 1800] loss 328.1488 l_rec 162.1160 l_s 553.4426 l_aux 0.0000 l1x 1.00 dead 0/8192
401
+ [epoch 75 step 1900] loss 323.5029 l_rec 159.9205 l_s 545.2748 l_aux 0.0000 l1x 1.00 dead 0/8192
402
+ [epoch 79 step 2000] loss 323.9691 l_rec 158.7100 l_s 550.8638 l_aux 0.0000 l1x 1.00 dead 0/8192
403
+ [epoch 83 step 2100] loss 323.9309 l_rec 157.7371 l_s 553.9792 l_aux 0.0000 l1x 1.00 dead 0/8192
404
+ [epoch 87 step 2200] loss 316.7346 l_rec 154.8299 l_s 539.6821 l_aux 0.0000 l1x 1.00 dead 0/8192
405
+ [epoch 91 step 2300] loss 319.8101 l_rec 155.4171 l_s 547.9767 l_aux 0.0000 l1x 1.00 dead 0/8192
406
+ [epoch 95 step 2400] loss 312.4319 l_rec 152.7612 l_s 532.2356 l_aux 0.0000 l1x 1.00 dead 0/8192
407
+ [epoch 99 step 2500] loss 313.6340 l_rec 153.1484 l_s 534.9520 l_aux 0.0000 l1x 1.00 dead 0/8192
408
+ [step 3000] EV=0.7993 L0=972.4 (38.5s)
409
+ [step 3000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step3000.pt
410
+ [step 4000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
411
+ [step 4000] preprocess center_scale: mean-shift=0.229 scale=0.0432
412
+ [epoch 3 step 100] loss 740.5062 l_rec 740.0449 l_s 3.8832 l_aux 0.0000 l1x 0.40 dead 0/8192
413
+ [epoch 7 step 200] loss 527.0967 l_rec 466.3249 l_s 254.4886 l_aux 0.0000 l1x 0.80 dead 0/8192
414
+ [epoch 11 step 300] loss 429.8510 l_rec 309.8146 l_s 400.1214 l_aux 0.0000 l1x 1.00 dead 0/8192
415
+ [epoch 15 step 400] loss 390.2275 l_rec 253.2363 l_s 456.6375 l_aux 0.0000 l1x 1.00 dead 0/8192
416
+ [epoch 19 step 500] loss 373.5827 l_rec 224.2609 l_s 497.7393 l_aux 0.0000 l1x 1.00 dead 0/8192
417
+ [epoch 23 step 600] loss 368.7617 l_rec 210.3364 l_s 528.0841 l_aux 0.0000 l1x 1.00 dead 0/8192
418
+ [epoch 27 step 700] loss 367.8697 l_rec 201.9138 l_s 553.1863 l_aux 0.0000 l1x 1.00 dead 0/8192
419
+ [epoch 31 step 800] loss 363.4539 l_rec 195.5704 l_s 559.6118 l_aux 0.0000 l1x 1.00 dead 0/8192
420
+ [epoch 35 step 900] loss 352.4101 l_rec 187.4891 l_s 549.7368 l_aux 0.0000 l1x 1.00 dead 0/8192
421
+ [epoch 39 step 1000] loss 359.1315 l_rec 186.3377 l_s 575.9792 l_aux 0.0000 l1x 1.00 dead 0/8192
422
+ [epoch 43 step 1100] loss 346.2921 l_rec 179.6142 l_s 555.5928 l_aux 0.0000 l1x 1.00 dead 0/8192
423
+ [epoch 47 step 1200] loss 343.1327 l_rec 176.8264 l_s 554.3545 l_aux 0.0000 l1x 1.00 dead 0/8192
424
+ [epoch 51 step 1300] loss 342.4291 l_rec 174.0688 l_s 561.2011 l_aux 0.0000 l1x 1.00 dead 0/8192
425
+ [epoch 55 step 1400] loss 344.6682 l_rec 172.6550 l_s 573.3772 l_aux 0.0000 l1x 1.00 dead 0/8192
426
+ [epoch 59 step 1500] loss 339.4234 l_rec 169.2697 l_s 567.1790 l_aux 0.0000 l1x 1.00 dead 0/8192
427
+ [epoch 63 step 1600] loss 336.5468 l_rec 167.0400 l_s 565.0226 l_aux 0.0000 l1x 1.00 dead 0/8192
428
+ [epoch 67 step 1700] loss 335.0789 l_rec 165.1505 l_s 566.4278 l_aux 0.0000 l1x 1.00 dead 0/8192
429
+ [epoch 71 step 1800] loss 333.4050 l_rec 163.5466 l_s 566.1948 l_aux 0.0000 l1x 1.00 dead 0/8192
430
+ [epoch 75 step 1900] loss 328.1611 l_rec 160.9946 l_s 557.2216 l_aux 0.0000 l1x 1.00 dead 0/8192
431
+ [epoch 79 step 2000] loss 328.3428 l_rec 159.9215 l_s 561.4042 l_aux 0.0000 l1x 1.00 dead 0/8192
432
+ [epoch 83 step 2100] loss 327.6591 l_rec 158.9055 l_s 562.5120 l_aux 0.0000 l1x 1.00 dead 0/8192
433
+ [epoch 87 step 2200] loss 321.1862 l_rec 156.0052 l_s 550.6033 l_aux 0.0000 l1x 1.00 dead 0/8192
434
+ [epoch 91 step 2300] loss 324.5598 l_rec 156.6249 l_s 559.7831 l_aux 0.0000 l1x 1.00 dead 0/8192
435
+ [epoch 95 step 2400] loss 316.6108 l_rec 154.0556 l_s 541.8508 l_aux 0.0000 l1x 1.00 dead 0/8192
436
+ [epoch 99 step 2500] loss 318.3661 l_rec 154.2071 l_s 547.1969 l_aux 0.0000 l1x 1.00 dead 0/8192
437
+ [step 4000] EV=0.7978 L0=977.4 (38.6s)
438
+ [step 4000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step4000.pt
439
+ [step 5000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
440
+ [step 5000] preprocess center_scale: mean-shift=0.221 scale=0.0439
441
+ [epoch 3 step 100] loss 742.5182 l_rec 742.0952 l_s 3.5614 l_aux 0.0000 l1x 0.40 dead 0/8192
442
+ [epoch 7 step 200] loss 536.4402 l_rec 481.1273 l_s 231.6283 l_aux 0.0000 l1x 0.80 dead 0/8192
443
+ [epoch 11 step 300] loss 435.1243 l_rec 314.1258 l_s 403.3283 l_aux 0.0000 l1x 1.00 dead 0/8192
444
+ [epoch 15 step 400] loss 394.5012 l_rec 253.7828 l_s 469.0613 l_aux 0.0000 l1x 1.00 dead 0/8192
445
+ [epoch 19 step 500] loss 378.4962 l_rec 224.6174 l_s 512.9293 l_aux 0.0000 l1x 1.00 dead 0/8192
446
+ [epoch 23 step 600] loss 374.3762 l_rec 211.2679 l_s 543.6945 l_aux 0.0000 l1x 1.00 dead 0/8192
447
+ [epoch 27 step 700] loss 373.2513 l_rec 202.9558 l_s 567.6519 l_aux 0.0000 l1x 1.00 dead 0/8192
448
+ [epoch 31 step 800] loss 368.2692 l_rec 196.5368 l_s 572.4412 l_aux 0.0000 l1x 1.00 dead 0/8192
449
+ [epoch 35 step 900] loss 356.9845 l_rec 188.7485 l_s 560.7867 l_aux 0.0000 l1x 1.00 dead 0/8192
450
+ [epoch 39 step 1000] loss 364.1912 l_rec 187.4996 l_s 588.9719 l_aux 0.0000 l1x 1.00 dead 0/8192
451
+ [epoch 43 step 1100] loss 351.0651 l_rec 180.8663 l_s 567.3295 l_aux 0.0000 l1x 1.00 dead 0/8192
452
+ [epoch 47 step 1200] loss 347.0239 l_rec 177.6438 l_s 564.6005 l_aux 0.0000 l1x 1.00 dead 0/8192
453
+ [epoch 51 step 1300] loss 347.1268 l_rec 175.5099 l_s 572.0564 l_aux 0.0000 l1x 1.00 dead 0/8192
454
+ [epoch 55 step 1400] loss 348.9342 l_rec 173.5633 l_s 584.5698 l_aux 0.0000 l1x 1.00 dead 0/8192
455
+ [epoch 59 step 1500] loss 343.9992 l_rec 170.4216 l_s 578.5922 l_aux 0.0000 l1x 1.00 dead 0/8192
456
+ [epoch 63 step 1600] loss 341.0271 l_rec 168.2967 l_s 575.7679 l_aux 0.0000 l1x 1.00 dead 0/8192
457
+ [epoch 67 step 1700] loss 338.8427 l_rec 166.2122 l_s 575.4348 l_aux 0.0000 l1x 1.00 dead 0/8192
458
+ [epoch 71 step 1800] loss 337.2959 l_rec 164.3638 l_s 576.4404 l_aux 0.0000 l1x 1.00 dead 0/8192
459
+ [epoch 75 step 1900] loss 332.3921 l_rec 162.2850 l_s 567.0237 l_aux 0.0000 l1x 1.00 dead 0/8192
460
+ [epoch 79 step 2000] loss 331.6189 l_rec 161.1781 l_s 568.1359 l_aux 0.0000 l1x 1.00 dead 0/8192
461
+ [epoch 83 step 2100] loss 331.5577 l_rec 160.0965 l_s 571.5374 l_aux 0.0000 l1x 1.00 dead 0/8192
462
+ [epoch 87 step 2200] loss 325.4994 l_rec 157.1339 l_s 561.2183 l_aux 0.0000 l1x 1.00 dead 0/8192
463
+ [epoch 91 step 2300] loss 328.5510 l_rec 157.9028 l_s 568.8273 l_aux 0.0000 l1x 1.00 dead 0/8192
464
+ [epoch 95 step 2400] loss 320.7331 l_rec 155.1433 l_s 551.9659 l_aux 0.0000 l1x 1.00 dead 0/8192
465
+ [epoch 99 step 2500] loss 321.9036 l_rec 155.5798 l_s 554.4128 l_aux 0.0000 l1x 1.00 dead 0/8192
466
+ [step 5000] EV=0.7961 L0=981.8 (39.4s)
467
+ [step 5000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step5000.pt
468
+ [step 6000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
469
+ [step 6000] preprocess center_scale: mean-shift=0.211 scale=0.0444
470
+ [epoch 3 step 100] loss 744.4907 l_rec 744.0983 l_s 3.3023 l_aux 0.0000 l1x 0.40 dead 0/8192
471
+ [epoch 7 step 200] loss 543.4768 l_rec 491.9855 l_s 215.6253 l_aux 0.0000 l1x 0.80 dead 0/8192
472
+ [epoch 11 step 300] loss 438.6403 l_rec 316.7071 l_s 406.4440 l_aux 0.0000 l1x 1.00 dead 0/8192
473
+ [epoch 15 step 400] loss 398.1854 l_rec 254.1791 l_s 480.0209 l_aux 0.0000 l1x 1.00 dead 0/8192
474
+ [epoch 19 step 500] loss 382.6852 l_rec 224.9142 l_s 525.9034 l_aux 0.0000 l1x 1.00 dead 0/8192
475
+ [epoch 23 step 600] loss 378.4837 l_rec 211.5930 l_s 556.3023 l_aux 0.0000 l1x 1.00 dead 0/8192
476
+ [epoch 27 step 700] loss 377.2396 l_rec 203.7099 l_s 578.4320 l_aux 0.0000 l1x 1.00 dead 0/8192
477
+ [epoch 31 step 800] loss 372.1085 l_rec 197.4173 l_s 582.3040 l_aux 0.0000 l1x 1.00 dead 0/8192
478
+ [epoch 35 step 900] loss 361.3077 l_rec 189.7150 l_s 571.9757 l_aux 0.0000 l1x 1.00 dead 0/8192
479
+ [epoch 39 step 1000] loss 367.5324 l_rec 188.3948 l_s 597.1255 l_aux 0.0000 l1x 1.00 dead 0/8192
480
+ [epoch 43 step 1100] loss 354.8302 l_rec 181.7803 l_s 576.8332 l_aux 0.0000 l1x 1.00 dead 0/8192
481
+ [epoch 47 step 1200] loss 350.3335 l_rec 178.4899 l_s 572.8118 l_aux 0.0000 l1x 1.00 dead 0/8192
482
+ [epoch 51 step 1300] loss 350.6716 l_rec 176.4226 l_s 580.8301 l_aux 0.0000 l1x 1.00 dead 0/8192
483
+ [epoch 55 step 1400] loss 351.6135 l_rec 174.3136 l_s 590.9996 l_aux 0.0000 l1x 1.00 dead 0/8192
484
+ [epoch 59 step 1500] loss 347.2458 l_rec 171.4529 l_s 585.9764 l_aux 0.0000 l1x 1.00 dead 0/8192
485
+ [epoch 63 step 1600] loss 344.2526 l_rec 169.2822 l_s 583.2347 l_aux 0.0000 l1x 1.00 dead 0/8192
486
+ [epoch 67 step 1700] loss 341.9704 l_rec 167.1360 l_s 582.7814 l_aux 0.0000 l1x 1.00 dead 0/8192
487
+ [epoch 71 step 1800] loss 339.6870 l_rec 165.1947 l_s 581.6412 l_aux 0.0000 l1x 1.00 dead 0/8192
488
+ [epoch 75 step 1900] loss 335.8210 l_rec 163.1896 l_s 575.4380 l_aux 0.0000 l1x 1.00 dead 0/8192
489
+ [epoch 79 step 2000] loss 334.3777 l_rec 161.7407 l_s 575.4568 l_aux 0.0000 l1x 1.00 dead 0/8192
490
+ [epoch 83 step 2100] loss 334.2598 l_rec 161.0014 l_s 577.5279 l_aux 0.0000 l1x 1.00 dead 0/8192
491
+ [epoch 87 step 2200] loss 328.5699 l_rec 158.3004 l_s 567.5652 l_aux 0.0000 l1x 1.00 dead 0/8192
492
+ [epoch 91 step 2300] loss 330.8861 l_rec 158.7811 l_s 573.6835 l_aux 0.0000 l1x 1.00 dead 0/8192
493
+ [epoch 95 step 2400] loss 323.4182 l_rec 156.0299 l_s 557.9610 l_aux 0.0000 l1x 1.00 dead 0/8192
494
+ [epoch 99 step 2500] loss 325.1666 l_rec 156.6023 l_s 561.8810 l_aux 0.0000 l1x 1.00 dead 0/8192
495
+ [step 6000] EV=0.7948 L0=985.5 (38.6s)
496
+ [step 6000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step6000.pt
497
+ [step 7000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
498
+ [step 7000] preprocess center_scale: mean-shift=0.208 scale=0.0448
499
+ [epoch 3 step 100] loss 746.8724 l_rec 746.4969 l_s 3.1608 l_aux 0.0000 l1x 0.40 dead 0/8192
500
+ [epoch 7 step 200] loss 548.3019 l_rec 499.7880 l_s 203.1569 l_aux 0.0000 l1x 0.80 dead 0/8192
501
+ [epoch 11 step 300] loss 440.5608 l_rec 317.9845 l_s 408.5877 l_aux 0.0000 l1x 1.00 dead 0/8192
502
+ [epoch 15 step 400] loss 400.5067 l_rec 254.0880 l_s 488.0623 l_aux 0.0000 l1x 1.00 dead 0/8192
503
+ [epoch 19 step 500] loss 385.5126 l_rec 224.7175 l_s 535.9839 l_aux 0.0000 l1x 1.00 dead 0/8192
504
+ [epoch 23 step 600] loss 381.5607 l_rec 212.0520 l_s 565.0289 l_aux 0.0000 l1x 1.00 dead 0/8192
505
+ [epoch 27 step 700] loss 380.3450 l_rec 204.3453 l_s 586.6658 l_aux 0.0000 l1x 1.00 dead 0/8192
506
+ [epoch 31 step 800] loss 375.0977 l_rec 198.0686 l_s 590.0969 l_aux 0.0000 l1x 1.00 dead 0/8192
507
+ [epoch 35 step 900] loss 364.1536 l_rec 190.5029 l_s 578.8357 l_aux 0.0000 l1x 1.00 dead 0/8192
508
+ [epoch 39 step 1000] loss 369.9399 l_rec 189.0297 l_s 603.0339 l_aux 0.0000 l1x 1.00 dead 0/8192
509
+ [epoch 43 step 1100] loss 357.6412 l_rec 182.4226 l_s 584.0621 l_aux 0.0000 l1x 1.00 dead 0/8192
510
+ [epoch 47 step 1200] loss 352.3735 l_rec 178.8321 l_s 578.4713 l_aux 0.0000 l1x 1.00 dead 0/8192
511
+ [epoch 51 step 1300] loss 353.4826 l_rec 177.2217 l_s 587.5363 l_aux 0.0000 l1x 1.00 dead 0/8192
512
+ [epoch 55 step 1400] loss 353.9917 l_rec 174.9958 l_s 596.6530 l_aux 0.0000 l1x 1.00 dead 0/8192
513
+ [epoch 59 step 1500] loss 350.1165 l_rec 172.3229 l_s 592.6453 l_aux 0.0000 l1x 1.00 dead 0/8192
514
+ [epoch 63 step 1600] loss 346.7851 l_rec 170.0164 l_s 589.2290 l_aux 0.0000 l1x 1.00 dead 0/8192
515
+ [epoch 67 step 1700] loss 343.7538 l_rec 167.6944 l_s 586.8647 l_aux 0.0000 l1x 1.00 dead 0/8192
516
+ [epoch 71 step 1800] loss 341.5882 l_rec 165.9400 l_s 585.4940 l_aux 0.0000 l1x 1.00 dead 0/8192
517
+ [epoch 75 step 1900] loss 338.2888 l_rec 164.0516 l_s 580.7906 l_aux 0.0000 l1x 1.00 dead 0/8192
518
+ [epoch 79 step 2000] loss 336.4110 l_rec 162.4290 l_s 579.9399 l_aux 0.0000 l1x 1.00 dead 0/8192
519
+ [epoch 83 step 2100] loss 335.8883 l_rec 161.6331 l_s 580.8505 l_aux 0.0000 l1x 1.00 dead 0/8192
520
+ [epoch 87 step 2200] loss 330.4790 l_rec 159.0255 l_s 571.5116 l_aux 0.0000 l1x 1.00 dead 0/8192
521
+ [epoch 91 step 2300] loss 333.1871 l_rec 159.6630 l_s 578.4135 l_aux 0.0000 l1x 1.00 dead 0/8192
522
+ [epoch 95 step 2400] loss 325.5021 l_rec 156.9439 l_s 561.8607 l_aux 0.0000 l1x 1.00 dead 0/8192
523
+ [epoch 99 step 2500] loss 327.1316 l_rec 157.2296 l_s 566.3400 l_aux 0.0000 l1x 1.00 dead 0/8192
524
+ [step 7000] EV=0.7938 L0=988.1 (38.4s)
525
+ [step 7000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step7000.pt
526
+ [step 8000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
527
+ [step 8000] preprocess center_scale: mean-shift=0.212 scale=0.0451
528
+ [epoch 3 step 100] loss 748.0084 l_rec 747.6494 l_s 3.0215 l_aux 0.0000 l1x 0.40 dead 0/8192
529
+ [epoch 7 step 200] loss 551.5128 l_rec 505.4889 l_s 192.7299 l_aux 0.0000 l1x 0.80 dead 0/8192
530
+ [epoch 11 step 300] loss 441.7218 l_rec 318.2473 l_s 411.5816 l_aux 0.0000 l1x 1.00 dead 0/8192
531
+ [epoch 15 step 400] loss 402.5273 l_rec 253.7971 l_s 495.7671 l_aux 0.0000 l1x 1.00 dead 0/8192
532
+ [epoch 19 step 500] loss 387.7733 l_rec 224.6097 l_s 543.8788 l_aux 0.0000 l1x 1.00 dead 0/8192
533
+ [epoch 23 step 600] loss 383.9175 l_rec 212.3727 l_s 571.8159 l_aux 0.0000 l1x 1.00 dead 0/8192
534
+ [epoch 27 step 700] loss 382.6304 l_rec 204.7571 l_s 592.9109 l_aux 0.0000 l1x 1.00 dead 0/8192
535
+ [epoch 31 step 800] loss 377.2741 l_rec 198.4498 l_s 596.0811 l_aux 0.0000 l1x 1.00 dead 0/8192
536
+ [epoch 35 step 900] loss 366.6888 l_rec 191.2541 l_s 584.7825 l_aux 0.0000 l1x 1.00 dead 0/8192
537
+ [epoch 39 step 1000] loss 371.5455 l_rec 189.3878 l_s 607.1922 l_aux 0.0000 l1x 1.00 dead 0/8192
538
+ [epoch 43 step 1100] loss 359.3738 l_rec 182.9758 l_s 587.9933 l_aux 0.0000 l1x 1.00 dead 0/8192
539
+ [epoch 47 step 1200] loss 353.8040 l_rec 179.2515 l_s 581.8419 l_aux 0.0000 l1x 1.00 dead 0/8192
540
+ [epoch 51 step 1300] loss 355.6597 l_rec 177.7389 l_s 593.0695 l_aux 0.0000 l1x 1.00 dead 0/8192
541
+ [epoch 55 step 1400] loss 355.8033 l_rec 175.4899 l_s 601.0447 l_aux 0.0000 l1x 1.00 dead 0/8192
542
+ [epoch 59 step 1500] loss 351.8857 l_rec 172.8725 l_s 596.7109 l_aux 0.0000 l1x 1.00 dead 0/8192
543
+ [epoch 63 step 1600] loss 348.3666 l_rec 170.4025 l_s 593.2136 l_aux 0.0000 l1x 1.00 dead 0/8192
544
+ [epoch 67 step 1700] loss 345.1306 l_rec 168.1080 l_s 590.0754 l_aux 0.0000 l1x 1.00 dead 0/8192
545
+ [epoch 71 step 1800] loss 343.3078 l_rec 166.4787 l_s 589.4305 l_aux 0.0000 l1x 1.00 dead 0/8192
546
+ [epoch 75 step 1900] loss 339.9758 l_rec 164.5485 l_s 584.7576 l_aux 0.0000 l1x 1.00 dead 0/8192
547
+ [epoch 79 step 2000] loss 337.5134 l_rec 162.9145 l_s 581.9962 l_aux 0.0000 l1x 1.00 dead 0/8192
548
+ [epoch 83 step 2100] loss 337.3812 l_rec 162.1837 l_s 583.9918 l_aux 0.0000 l1x 1.00 dead 0/8192
549
+ [epoch 87 step 2200] loss 332.0016 l_rec 159.4539 l_s 575.1589 l_aux 0.0000 l1x 1.00 dead 0/8192
550
+ [epoch 91 step 2300] loss 334.7272 l_rec 160.3075 l_s 581.3989 l_aux 0.0000 l1x 1.00 dead 0/8192
551
+ [epoch 95 step 2400] loss 327.0352 l_rec 157.6196 l_s 564.7186 l_aux 0.0000 l1x 1.00 dead 0/8192
552
+ [epoch 99 step 2500] loss 328.7789 l_rec 157.7952 l_s 569.9457 l_aux 0.0000 l1x 1.00 dead 0/8192
553
+ [step 8000] EV=0.7931 L0=990.1 (38.7s)
554
+ [step 8000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step8000.pt
555
+ [step 9000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
556
+ [step 9000] preprocess center_scale: mean-shift=0.214 scale=0.0454
557
+ [epoch 3 step 100] loss 748.7295 l_rec 748.3830 l_s 2.9169 l_aux 0.0000 l1x 0.40 dead 0/8192
558
+ [epoch 7 step 200] loss 555.1830 l_rec 511.2586 l_s 183.9380 l_aux 0.0000 l1x 0.80 dead 0/8192
559
+ [epoch 11 step 300] loss 442.8305 l_rec 318.2896 l_s 415.1365 l_aux 0.0000 l1x 1.00 dead 0/8192
560
+ [epoch 15 step 400] loss 403.6893 l_rec 252.8535 l_s 502.7858 l_aux 0.0000 l1x 1.00 dead 0/8192
561
+ [epoch 19 step 500] loss 389.5502 l_rec 224.2678 l_s 550.9413 l_aux 0.0000 l1x 1.00 dead 0/8192
562
+ [epoch 23 step 600] loss 385.9750 l_rec 212.4354 l_s 578.4656 l_aux 0.0000 l1x 1.00 dead 0/8192
563
+ [epoch 27 step 700] loss 384.1175 l_rec 204.8409 l_s 597.5889 l_aux 0.0000 l1x 1.00 dead 0/8192
564
+ [epoch 31 step 800] loss 378.9205 l_rec 198.7980 l_s 600.4084 l_aux 0.0000 l1x 1.00 dead 0/8192
565
+ [epoch 35 step 900] loss 367.9637 l_rec 191.4410 l_s 588.4089 l_aux 0.0000 l1x 1.00 dead 0/8192
566
+ [epoch 39 step 1000] loss 372.6650 l_rec 189.6879 l_s 609.9237 l_aux 0.0000 l1x 1.00 dead 0/8192
567
+ [epoch 43 step 1100] loss 360.7657 l_rec 183.3032 l_s 591.5418 l_aux 0.0000 l1x 1.00 dead 0/8192
568
+ [epoch 47 step 1200] loss 355.0594 l_rec 179.5348 l_s 585.0819 l_aux 0.0000 l1x 1.00 dead 0/8192
569
+ [epoch 51 step 1300] loss 357.4457 l_rec 178.2321 l_s 597.3788 l_aux 0.0000 l1x 1.00 dead 0/8192
570
+ [epoch 55 step 1400] loss 357.0114 l_rec 175.9285 l_s 603.6096 l_aux 0.0000 l1x 1.00 dead 0/8192
571
+ [epoch 59 step 1500] loss 353.1621 l_rec 173.4889 l_s 598.9105 l_aux 0.0000 l1x 1.00 dead 0/8192
572
+ [epoch 63 step 1600] loss 349.8860 l_rec 171.0695 l_s 596.0549 l_aux 0.0000 l1x 1.00 dead 0/8192
573
+ [epoch 67 step 1700] loss 346.3940 l_rec 168.6274 l_s 592.5553 l_aux 0.0000 l1x 1.00 dead 0/8192
574
+ [epoch 71 step 1800] loss 344.3709 l_rec 167.0431 l_s 591.0927 l_aux 0.0000 l1x 1.00 dead 0/8192
575
+ [epoch 75 step 1900] loss 341.4211 l_rec 165.1655 l_s 587.5189 l_aux 0.0000 l1x 1.00 dead 0/8192
576
+ [epoch 79 step 2000] loss 338.7019 l_rec 163.4102 l_s 584.3056 l_aux 0.0000 l1x 1.00 dead 0/8192
577
+ [epoch 83 step 2100] loss 338.9210 l_rec 162.6192 l_s 587.6724 l_aux 0.0000 l1x 1.00 dead 0/8192
578
+ [epoch 87 step 2200] loss 333.3380 l_rec 159.9749 l_s 577.8770 l_aux 0.0000 l1x 1.00 dead 0/8192
579
+ [epoch 91 step 2300] loss 336.0588 l_rec 160.9279 l_s 583.7697 l_aux 0.0000 l1x 1.00 dead 0/8192
580
+ [epoch 95 step 2400] loss 328.7147 l_rec 158.0122 l_s 569.0083 l_aux 0.0000 l1x 1.00 dead 0/8192
581
+ [epoch 99 step 2500] loss 330.2180 l_rec 158.2281 l_s 573.2998 l_aux 0.0000 l1x 1.00 dead 0/8192
582
+ [step 9000] EV=0.7924 L0=992.0 (38.3s)
583
+ [step 9000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step9000.pt
584
+ [step 14000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
585
+ [step 14000] preprocess center_scale: mean-shift=0.240 scale=0.0459
586
+ [epoch 3 step 100] loss 751.0924 l_rec 750.7834 l_s 2.6006 l_aux 0.0000 l1x 0.40 dead 0/8192
587
+ [epoch 7 step 200] loss 565.9417 l_rec 531.5352 l_s 144.0811 l_aux 0.0000 l1x 0.80 dead 0/8192
588
+ [epoch 11 step 300] loss 443.6217 l_rec 312.6878 l_s 436.4465 l_aux 0.0000 l1x 1.00 dead 0/8192
589
+ [epoch 15 step 400] loss 406.1838 l_rec 248.7146 l_s 524.8975 l_aux 0.0000 l1x 1.00 dead 0/8192
590
+ [epoch 19 step 500] loss 394.2518 l_rec 223.1194 l_s 570.4413 l_aux 0.0000 l1x 1.00 dead 0/8192
591
+ [epoch 23 step 600] loss 390.2458 l_rec 212.0677 l_s 593.9269 l_aux 0.0000 l1x 1.00 dead 0/8192
592
+ [epoch 27 step 700] loss 388.9763 l_rec 205.6408 l_s 611.1183 l_aux 0.0000 l1x 1.00 dead 0/8192
593
+ [epoch 31 step 800] loss 383.2409 l_rec 199.3953 l_s 612.8188 l_aux 0.0000 l1x 1.00 dead 0/8192
594
+ [epoch 35 step 900] loss 372.6218 l_rec 192.4213 l_s 600.6683 l_aux 0.0000 l1x 1.00 dead 0/8192
595
+ [epoch 39 step 1000] loss 376.1484 l_rec 190.4080 l_s 619.1348 l_aux 0.0000 l1x 1.00 dead 0/8192
596
+ [epoch 43 step 1100] loss 364.8875 l_rec 183.9796 l_s 603.0262 l_aux 0.0000 l1x 1.00 dead 0/8192
597
+ [epoch 47 step 1200] loss 357.7644 l_rec 179.9999 l_s 592.5483 l_aux 0.0000 l1x 1.00 dead 0/8192
598
+ [epoch 51 step 1300] loss 361.2602 l_rec 179.0752 l_s 607.2834 l_aux 0.0000 l1x 1.00 dead 0/8192
599
+ [epoch 55 step 1400] loss 359.8582 l_rec 176.3061 l_s 611.8401 l_aux 0.0000 l1x 1.00 dead 0/8192
600
+ [epoch 59 step 1500] loss 357.0620 l_rec 174.3875 l_s 608.9149 l_aux 0.0000 l1x 1.00 dead 0/8192
601
+ [epoch 63 step 1600] loss 354.1461 l_rec 172.1171 l_s 606.7634 l_aux 0.0000 l1x 1.00 dead 0/8192
602
+ [epoch 67 step 1700] loss 349.6859 l_rec 169.5039 l_s 600.6068 l_aux 0.0000 l1x 1.00 dead 0/8192
603
+ [epoch 71 step 1800] loss 347.1339 l_rec 167.7590 l_s 597.9163 l_aux 0.0000 l1x 1.00 dead 0/8192
604
+ [epoch 75 step 1900] loss 345.0735 l_rec 165.9642 l_s 597.0312 l_aux 0.0000 l1x 1.00 dead 0/8192
605
+ [epoch 79 step 2000] loss 342.0947 l_rec 164.1150 l_s 593.2656 l_aux 0.0000 l1x 1.00 dead 0/8192
606
+ [epoch 83 step 2100] loss 341.4760 l_rec 163.5302 l_s 593.1527 l_aux 0.0000 l1x 1.00 dead 0/8192
607
+ [epoch 87 step 2200] loss 336.2534 l_rec 161.1172 l_s 583.7875 l_aux 0.0000 l1x 1.00 dead 0/8192
608
+ [epoch 91 step 2300] loss 339.5181 l_rec 161.8795 l_s 592.1286 l_aux 0.0000 l1x 1.00 dead 0/8192
609
+ [epoch 95 step 2400] loss 331.6540 l_rec 158.9054 l_s 575.8286 l_aux 0.0000 l1x 1.00 dead 0/8192
610
+ [epoch 99 step 2500] loss 333.4131 l_rec 159.4804 l_s 579.7758 l_aux 0.0000 l1x 1.00 dead 0/8192
611
+ [step 14000] EV=0.7912 L0=996.5 (38.5s)
612
+ [step 14000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step14000.pt
613
+ [step 21000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
614
+ [step 21000] preprocess center_scale: mean-shift=0.261 scale=0.0457
615
+ [epoch 3 step 100] loss 752.7520 l_rec 752.4691 l_s 2.3819 l_aux 0.0000 l1x 0.40 dead 0/8192
616
+ [epoch 7 step 200] loss 575.2158 l_rec 547.0101 l_s 118.1145 l_aux 0.0000 l1x 0.80 dead 0/8192
617
+ [epoch 11 step 300] loss 442.8688 l_rec 307.3776 l_s 451.6373 l_aux 0.0000 l1x 1.00 dead 0/8192
618
+ [epoch 15 step 400] loss 407.3076 l_rec 245.0149 l_s 540.9757 l_aux 0.0000 l1x 1.00 dead 0/8192
619
+ [epoch 19 step 500] loss 396.5739 l_rec 221.7607 l_s 582.7106 l_aux 0.0000 l1x 1.00 dead 0/8192
620
+ [epoch 23 step 600] loss 393.1835 l_rec 211.8372 l_s 604.4877 l_aux 0.0000 l1x 1.00 dead 0/8192
621
+ [epoch 27 step 700] loss 391.7235 l_rec 205.5830 l_s 620.4684 l_aux 0.0000 l1x 1.00 dead 0/8192
622
+ [epoch 31 step 800] loss 385.6569 l_rec 199.5587 l_s 620.3274 l_aux 0.0000 l1x 1.00 dead 0/8192
623
+ [epoch 35 step 900] loss 374.9463 l_rec 192.8884 l_s 606.8598 l_aux 0.0000 l1x 1.00 dead 0/8192
624
+ [epoch 39 step 1000] loss 377.3926 l_rec 190.5412 l_s 622.8380 l_aux 0.0000 l1x 1.00 dead 0/8192
625
+ [epoch 43 step 1100] loss 366.3987 l_rec 184.2560 l_s 607.1425 l_aux 0.0000 l1x 1.00 dead 0/8192
626
+ [epoch 47 step 1200] loss 359.5332 l_rec 180.3353 l_s 597.3261 l_aux 0.0000 l1x 1.00 dead 0/8192
627
+ [epoch 51 step 1300] loss 363.1786 l_rec 179.4845 l_s 612.3137 l_aux 0.0000 l1x 1.00 dead 0/8192
628
+ [epoch 55 step 1400] loss 361.2748 l_rec 176.5169 l_s 615.8596 l_aux 0.0000 l1x 1.00 dead 0/8192
629
+ [epoch 59 step 1500] loss 359.0342 l_rec 174.6983 l_s 614.4530 l_aux 0.0000 l1x 1.00 dead 0/8192
630
+ [epoch 63 step 1600] loss 356.1245 l_rec 172.5867 l_s 611.7929 l_aux 0.0000 l1x 1.00 dead 0/8192
631
+ [epoch 67 step 1700] loss 350.5089 l_rec 169.5514 l_s 603.1916 l_aux 0.0000 l1x 1.00 dead 0/8192
632
+ [epoch 71 step 1800] loss 349.1537 l_rec 168.1255 l_s 603.4272 l_aux 0.0000 l1x 1.00 dead 0/8192
633
+ [epoch 75 step 1900] loss 346.4178 l_rec 166.2610 l_s 600.5225 l_aux 0.0000 l1x 1.00 dead 0/8192
634
+ [epoch 79 step 2000] loss 343.8334 l_rec 164.2214 l_s 598.7067 l_aux 0.0000 l1x 1.00 dead 0/8192
635
+ [epoch 83 step 2100] loss 342.8062 l_rec 163.8851 l_s 596.4033 l_aux 0.0000 l1x 1.00 dead 0/8192
636
+ [epoch 87 step 2200] loss 337.9162 l_rec 161.4527 l_s 588.2114 l_aux 0.0000 l1x 1.00 dead 0/8192
637
+ [epoch 91 step 2300] loss 340.8118 l_rec 162.1959 l_s 595.3864 l_aux 0.0000 l1x 1.00 dead 0/8192
638
+ [epoch 95 step 2400] loss 332.5220 l_rec 159.2816 l_s 577.4680 l_aux 0.0000 l1x 1.00 dead 0/8192
639
+ [epoch 99 step 2500] loss 334.7898 l_rec 159.6807 l_s 583.6970 l_aux 0.0000 l1x 1.00 dead 0/8192
640
+ [step 21000] EV=0.7908 L0=998.5 (38.4s)
641
+ [step 21000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step21000.pt
642
+ [step 27000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
643
+ [step 27000] preprocess center_scale: mean-shift=0.268 scale=0.0451
644
+ [epoch 3 step 100] loss 753.4792 l_rec 753.2059 l_s 2.3006 l_aux 0.0000 l1x 0.40 dead 0/8192
645
+ [epoch 7 step 200] loss 580.2642 l_rec 554.4330 l_s 108.1707 l_aux 0.0000 l1x 0.80 dead 0/8192
646
+ [epoch 11 step 300] loss 442.5096 l_rec 304.0689 l_s 461.4688 l_aux 0.0000 l1x 1.00 dead 0/8192
647
+ [epoch 15 step 400] loss 408.1027 l_rec 243.9213 l_s 547.2714 l_aux 0.0000 l1x 1.00 dead 0/8192
648
+ [epoch 19 step 500] loss 397.3169 l_rec 221.1851 l_s 587.1058 l_aux 0.0000 l1x 1.00 dead 0/8192
649
+ [epoch 23 step 600] loss 394.4814 l_rec 211.8358 l_s 608.8185 l_aux 0.0000 l1x 1.00 dead 0/8192
650
+ [epoch 27 step 700] loss 392.6103 l_rec 205.5041 l_s 623.6875 l_aux 0.0000 l1x 1.00 dead 0/8192
651
+ [epoch 31 step 800] loss 386.9615 l_rec 199.7081 l_s 624.1779 l_aux 0.0000 l1x 1.00 dead 0/8192
652
+ [epoch 35 step 900] loss 375.5479 l_rec 192.7681 l_s 609.2661 l_aux 0.0000 l1x 1.00 dead 0/8192
653
+ [epoch 39 step 1000] loss 378.2921 l_rec 190.3303 l_s 626.5393 l_aux 0.0000 l1x 1.00 dead 0/8192
654
+ [epoch 43 step 1100] loss 366.8534 l_rec 184.3254 l_s 608.4266 l_aux 0.0000 l1x 1.00 dead 0/8192
655
+ [epoch 47 step 1200] loss 359.4331 l_rec 180.0723 l_s 597.8693 l_aux 0.0000 l1x 1.00 dead 0/8192
656
+ [epoch 51 step 1300] loss 363.5555 l_rec 179.4244 l_s 613.7703 l_aux 0.0000 l1x 1.00 dead 0/8192
657
+ [epoch 55 step 1400] loss 361.4332 l_rec 176.4534 l_s 616.5995 l_aux 0.0000 l1x 1.00 dead 0/8192
658
+ [epoch 59 step 1500] loss 359.5349 l_rec 174.8604 l_s 615.5815 l_aux 0.0000 l1x 1.00 dead 0/8192
659
+ [epoch 63 step 1600] loss 356.3924 l_rec 172.4239 l_s 613.2284 l_aux 0.0000 l1x 1.00 dead 0/8192
660
+ [epoch 67 step 1700] loss 351.0789 l_rec 169.5161 l_s 605.2094 l_aux 0.0000 l1x 1.00 dead 0/8192
661
+ [epoch 71 step 1800] loss 349.3942 l_rec 168.1728 l_s 604.0710 l_aux 0.0000 l1x 1.00 dead 0/8192
662
+ [epoch 75 step 1900] loss 346.8262 l_rec 166.2311 l_s 601.9836 l_aux 0.0000 l1x 1.00 dead 0/8192
663
+ [epoch 79 step 2000] loss 344.3245 l_rec 164.5431 l_s 599.2714 l_aux 0.0000 l1x 1.00 dead 0/8192
664
+ [epoch 83 step 2100] loss 343.4041 l_rec 164.0930 l_s 597.7035 l_aux 0.0000 l1x 1.00 dead 0/8192
665
+ [epoch 87 step 2200] loss 338.3742 l_rec 161.3891 l_s 589.9503 l_aux 0.0000 l1x 1.00 dead 0/8192
666
+ [epoch 91 step 2300] loss 341.8085 l_rec 162.3672 l_s 598.1376 l_aux 0.0000 l1x 1.00 dead 0/8192
667
+ [epoch 95 step 2400] loss 333.2529 l_rec 159.3019 l_s 579.8365 l_aux 0.0000 l1x 1.00 dead 0/8192
668
+ [epoch 99 step 2500] loss 335.1691 l_rec 159.6539 l_s 585.0504 l_aux 0.0000 l1x 1.00 dead 0/8192
669
+ [step 27000] EV=0.7908 L0=999.3 (39.2s)
670
+ [step 27000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step27000.pt
671
+ [step 34000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
672
+ [step 34000] preprocess center_scale: mean-shift=0.268 scale=0.0443
673
+ [epoch 3 step 100] loss 753.4775 l_rec 753.2070 l_s 2.2764 l_aux 0.0000 l1x 0.40 dead 0/8192
674
+ [epoch 7 step 200] loss 583.1448 l_rec 558.3083 l_s 104.0053 l_aux 0.0000 l1x 0.80 dead 0/8192
675
+ [epoch 11 step 300] loss 443.1728 l_rec 303.3618 l_s 466.0366 l_aux 0.0000 l1x 1.00 dead 0/8192
676
+ [epoch 15 step 400] loss 409.1614 l_rec 243.2718 l_s 552.9655 l_aux 0.0000 l1x 1.00 dead 0/8192
677
+ [epoch 19 step 500] loss 398.5327 l_rec 221.0751 l_s 591.5253 l_aux 0.0000 l1x 1.00 dead 0/8192
678
+ [epoch 23 step 600] loss 395.1291 l_rec 211.7722 l_s 611.1895 l_aux 0.0000 l1x 1.00 dead 0/8192
679
+ [epoch 27 step 700] loss 393.8130 l_rec 205.4430 l_s 627.9000 l_aux 0.0000 l1x 1.00 dead 0/8192
680
+ [epoch 31 step 800] loss 388.2473 l_rec 199.6077 l_s 628.7987 l_aux 0.0000 l1x 1.00 dead 0/8192
681
+ [epoch 35 step 900] loss 376.2881 l_rec 192.7553 l_s 611.7759 l_aux 0.0000 l1x 1.00 dead 0/8192
682
+ [epoch 39 step 1000] loss 378.8491 l_rec 190.4093 l_s 628.1325 l_aux 0.0000 l1x 1.00 dead 0/8192
683
+ [epoch 43 step 1100] loss 367.4835 l_rec 184.1544 l_s 611.0968 l_aux 0.0000 l1x 1.00 dead 0/8192
684
+ [epoch 47 step 1200] loss 360.2900 l_rec 180.1904 l_s 600.3319 l_aux 0.0000 l1x 1.00 dead 0/8192
685
+ [epoch 51 step 1300] loss 364.0220 l_rec 179.6270 l_s 614.6500 l_aux 0.0000 l1x 1.00 dead 0/8192
686
+ [epoch 55 step 1400] loss 361.9374 l_rec 176.5197 l_s 618.0589 l_aux 0.0000 l1x 1.00 dead 0/8192
687
+ [epoch 59 step 1500] loss 359.6822 l_rec 174.8347 l_s 616.1581 l_aux 0.0000 l1x 1.00 dead 0/8192
688
+ [epoch 63 step 1600] loss 357.4796 l_rec 172.6051 l_s 616.2484 l_aux 0.0000 l1x 1.00 dead 0/8192
689
+ [epoch 67 step 1700] loss 351.7040 l_rec 169.6898 l_s 606.7141 l_aux 0.0000 l1x 1.00 dead 0/8192
690
+ [epoch 71 step 1800] loss 350.1419 l_rec 168.1927 l_s 606.4974 l_aux 0.0000 l1x 1.00 dead 0/8192
691
+ [epoch 75 step 1900] loss 347.5459 l_rec 166.3732 l_s 603.9091 l_aux 0.0000 l1x 1.00 dead 0/8192
692
+ [epoch 79 step 2000] loss 344.6309 l_rec 164.5030 l_s 600.4264 l_aux 0.0000 l1x 1.00 dead 0/8192
693
+ [epoch 83 step 2100] loss 344.1322 l_rec 164.0521 l_s 600.2668 l_aux 0.0000 l1x 1.00 dead 0/8192
694
+ [epoch 87 step 2200] loss 338.7600 l_rec 161.5173 l_s 590.8090 l_aux 0.0000 l1x 1.00 dead 0/8192
695
+ [epoch 91 step 2300] loss 341.8137 l_rec 162.2291 l_s 598.6154 l_aux 0.0000 l1x 1.00 dead 0/8192
696
+ [epoch 95 step 2400] loss 333.5186 l_rec 159.5782 l_s 579.8014 l_aux 0.0000 l1x 1.00 dead 0/8192
697
+ [epoch 99 step 2500] loss 335.5377 l_rec 159.9650 l_s 585.2425 l_aux 0.0000 l1x 1.00 dead 0/8192
698
+ [step 34000] EV=0.7906 L0=999.5 (38.5s)
699
+ [step 34000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step34000.pt
700
+ [step 47000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
701
+ [step 47000] preprocess center_scale: mean-shift=0.250 scale=0.0422
702
+ [epoch 3 step 100] loss 752.9741 l_rec 752.7017 l_s 2.2927 l_aux 0.0000 l1x 0.40 dead 0/8192
703
+ [epoch 7 step 200] loss 585.2609 l_rec 559.3107 l_s 108.6691 l_aux 0.0000 l1x 0.80 dead 0/8192
704
+ [epoch 11 step 300] loss 446.7000 l_rec 307.0673 l_s 465.4422 l_aux 0.0000 l1x 1.00 dead 0/8192
705
+ [epoch 15 step 400] loss 410.7524 l_rec 243.5573 l_s 557.3169 l_aux 0.0000 l1x 1.00 dead 0/8192
706
+ [epoch 19 step 500] loss 399.7614 l_rec 221.4609 l_s 594.3350 l_aux 0.0000 l1x 1.00 dead 0/8192
707
+ [epoch 23 step 600] loss 395.7774 l_rec 211.7922 l_s 613.2841 l_aux 0.0000 l1x 1.00 dead 0/8192
708
+ [epoch 27 step 700] loss 394.4908 l_rec 205.5349 l_s 629.8529 l_aux 0.0000 l1x 1.00 dead 0/8192
709
+ [epoch 31 step 800] loss 389.3908 l_rec 200.0942 l_s 630.9888 l_aux 0.0000 l1x 1.00 dead 0/8192
710
+ [epoch 35 step 900] loss 377.3149 l_rec 192.9560 l_s 614.5294 l_aux 0.0000 l1x 1.00 dead 0/8192
711
+ [epoch 39 step 1000] loss 379.9300 l_rec 190.6675 l_s 630.8748 l_aux 0.0000 l1x 1.00 dead 0/8192
712
+ [epoch 43 step 1100] loss 368.3587 l_rec 184.5018 l_s 612.8561 l_aux 0.0000 l1x 1.00 dead 0/8192
713
+ [epoch 47 step 1200] loss 361.4451 l_rec 180.4276 l_s 603.3918 l_aux 0.0000 l1x 1.00 dead 0/8192
714
+ [epoch 51 step 1300] loss 364.5399 l_rec 179.7824 l_s 615.8581 l_aux 0.0000 l1x 1.00 dead 0/8192
715
+ [epoch 55 step 1400] loss 362.9869 l_rec 176.8835 l_s 620.3447 l_aux 0.0000 l1x 1.00 dead 0/8192
716
+ [epoch 59 step 1500] loss 360.2820 l_rec 174.9879 l_s 617.6470 l_aux 0.0000 l1x 1.00 dead 0/8192
717
+ [epoch 63 step 1600] loss 357.9425 l_rec 172.8268 l_s 617.0524 l_aux 0.0000 l1x 1.00 dead 0/8192
718
+ [epoch 67 step 1700] loss 351.9385 l_rec 169.6358 l_s 607.6755 l_aux 0.0000 l1x 1.00 dead 0/8192
719
+ [epoch 71 step 1800] loss 350.3969 l_rec 168.0163 l_s 607.9353 l_aux 0.0000 l1x 1.00 dead 0/8192
720
+ [epoch 75 step 1900] loss 347.8184 l_rec 166.2952 l_s 605.0773 l_aux 0.0000 l1x 1.00 dead 0/8192
721
+ [epoch 79 step 2000] loss 344.9127 l_rec 164.4869 l_s 601.4191 l_aux 0.0000 l1x 1.00 dead 0/8192
722
+ [epoch 83 step 2100] loss 344.5792 l_rec 163.9730 l_s 602.0207 l_aux 0.0000 l1x 1.00 dead 0/8192
723
+ [epoch 87 step 2200] loss 338.7360 l_rec 161.3034 l_s 591.4421 l_aux 0.0000 l1x 1.00 dead 0/8192
724
+ [epoch 91 step 2300] loss 341.9368 l_rec 162.0546 l_s 599.6074 l_aux 0.0000 l1x 1.00 dead 0/8192
725
+ [epoch 95 step 2400] loss 333.7814 l_rec 159.4305 l_s 581.1697 l_aux 0.0000 l1x 1.00 dead 0/8192
726
+ [epoch 99 step 2500] loss 336.3510 l_rec 159.7570 l_s 588.6468 l_aux 0.0000 l1x 1.00 dead 0/8192
727
+ [step 47000] EV=0.7906 L0=1000.3 (38.4s)
728
+ [step 47000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step47000.pt
729
+ [step 61000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
730
+ [step 61000] preprocess center_scale: mean-shift=0.236 scale=0.0393
731
+ [epoch 3 step 100] loss 752.7858 l_rec 752.4998 l_s 2.4068 l_aux 0.0000 l1x 0.40 dead 0/8192
732
+ [epoch 7 step 200] loss 582.9722 l_rec 552.6761 l_s 126.8683 l_aux 0.0000 l1x 0.80 dead 0/8192
733
+ [epoch 11 step 300] loss 452.3908 l_rec 316.3014 l_s 453.6312 l_aux 0.0000 l1x 1.00 dead 0/8192
734
+ [epoch 15 step 400] loss 412.7169 l_rec 245.0292 l_s 558.9591 l_aux 0.0000 l1x 1.00 dead 0/8192
735
+ [epoch 19 step 500] loss 401.1283 l_rec 222.1973 l_s 596.4365 l_aux 0.0000 l1x 1.00 dead 0/8192
736
+ [epoch 23 step 600] loss 396.7276 l_rec 212.2660 l_s 614.8721 l_aux 0.0000 l1x 1.00 dead 0/8192
737
+ [epoch 27 step 700] loss 395.8503 l_rec 205.8293 l_s 633.4033 l_aux 0.0000 l1x 1.00 dead 0/8192
738
+ [epoch 31 step 800] loss 390.0436 l_rec 200.2815 l_s 632.5403 l_aux 0.0000 l1x 1.00 dead 0/8192
739
+ [epoch 35 step 900] loss 378.4133 l_rec 193.4230 l_s 616.6340 l_aux 0.0000 l1x 1.00 dead 0/8192
740
+ [epoch 39 step 1000] loss 381.2427 l_rec 191.0609 l_s 633.9393 l_aux 0.0000 l1x 1.00 dead 0/8192
741
+ [epoch 43 step 1100] loss 368.8238 l_rec 184.7569 l_s 613.5563 l_aux 0.0000 l1x 1.00 dead 0/8192
742
+ [epoch 47 step 1200] loss 362.0282 l_rec 180.6933 l_s 604.4495 l_aux 0.0000 l1x 1.00 dead 0/8192
743
+ [epoch 51 step 1300] loss 365.1139 l_rec 179.7020 l_s 618.0396 l_aux 0.0000 l1x 1.00 dead 0/8192
744
+ [epoch 55 step 1400] loss 363.4353 l_rec 176.8927 l_s 621.8088 l_aux 0.0000 l1x 1.00 dead 0/8192
745
+ [epoch 59 step 1500] loss 361.7016 l_rec 175.1593 l_s 621.8077 l_aux 0.0000 l1x 1.00 dead 0/8192
746
+ [epoch 63 step 1600] loss 359.0020 l_rec 172.8757 l_s 620.4212 l_aux 0.0000 l1x 1.00 dead 0/8192
747
+ [epoch 67 step 1700] loss 352.1486 l_rec 169.7194 l_s 608.0974 l_aux 0.0000 l1x 1.00 dead 0/8192
748
+ [epoch 71 step 1800] loss 351.2077 l_rec 168.3088 l_s 609.6628 l_aux 0.0000 l1x 1.00 dead 0/8192
749
+ [epoch 75 step 1900] loss 348.7200 l_rec 166.5754 l_s 607.1486 l_aux 0.0000 l1x 1.00 dead 0/8192
750
+ [epoch 79 step 2000] loss 345.7929 l_rec 164.7476 l_s 603.4844 l_aux 0.0000 l1x 1.00 dead 0/8192
751
+ [epoch 83 step 2100] loss 345.3228 l_rec 164.2843 l_s 603.4616 l_aux 0.0000 l1x 1.00 dead 0/8192
752
+ [epoch 87 step 2200] loss 339.5664 l_rec 161.6288 l_s 593.1254 l_aux 0.0000 l1x 1.00 dead 0/8192
753
+ [epoch 91 step 2300] loss 342.9393 l_rec 162.2984 l_s 602.1362 l_aux 0.0000 l1x 1.00 dead 0/8192
754
+ [epoch 95 step 2400] loss 334.2091 l_rec 159.4351 l_s 582.5801 l_aux 0.0000 l1x 1.00 dead 0/8192
755
+ [epoch 99 step 2500] loss 336.8251 l_rec 160.0883 l_s 589.1227 l_aux 0.0000 l1x 1.00 dead 0/8192
756
+ [step 61000] EV=0.7904 L0=1001.8 (39.2s)
757
+ [step 61000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step61000.pt
758
+ [step 75000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
759
+ [step 75000] preprocess center_scale: mean-shift=0.224 scale=0.0361
760
+ [epoch 3 step 100] loss 750.7099 l_rec 750.4031 l_s 2.5820 l_aux 0.0000 l1x 0.40 dead 0/8192
761
+ [epoch 7 step 200] loss 578.6434 l_rec 540.8817 l_s 158.1315 l_aux 0.0000 l1x 0.80 dead 0/8192
762
+ [epoch 11 step 300] loss 459.4569 l_rec 328.1042 l_s 437.8422 l_aux 0.0000 l1x 1.00 dead 0/8192
763
+ [epoch 15 step 400] loss 414.2567 l_rec 248.0318 l_s 554.0829 l_aux 0.0000 l1x 1.00 dead 0/8192
764
+ [epoch 19 step 500] loss 402.2108 l_rec 223.1612 l_s 596.8319 l_aux 0.0000 l1x 1.00 dead 0/8192
765
+ [epoch 23 step 600] loss 398.3768 l_rec 213.0702 l_s 617.6887 l_aux 0.0000 l1x 1.00 dead 0/8192
766
+ [epoch 27 step 700] loss 397.2003 l_rec 206.4834 l_s 635.7231 l_aux 0.0000 l1x 1.00 dead 0/8192
767
+ [epoch 31 step 800] loss 391.4822 l_rec 200.9013 l_s 635.2697 l_aux 0.0000 l1x 1.00 dead 0/8192
768
+ [epoch 35 step 900] loss 380.1983 l_rec 194.1670 l_s 620.1044 l_aux 0.0000 l1x 1.00 dead 0/8192
769
+ [epoch 39 step 1000] loss 382.6296 l_rec 191.5113 l_s 637.0609 l_aux 0.0000 l1x 1.00 dead 0/8192
770
+ [epoch 43 step 1100] loss 369.9186 l_rec 185.0514 l_s 616.2240 l_aux 0.0000 l1x 1.00 dead 0/8192
771
+ [epoch 47 step 1200] loss 363.1601 l_rec 181.1069 l_s 606.8440 l_aux 0.0000 l1x 1.00 dead 0/8192
772
+ [epoch 51 step 1300] loss 366.0925 l_rec 179.9148 l_s 620.5925 l_aux 0.0000 l1x 1.00 dead 0/8192
773
+ [epoch 55 step 1400] loss 363.8982 l_rec 177.2237 l_s 622.2482 l_aux 0.0000 l1x 1.00 dead 0/8192
774
+ [epoch 59 step 1500] loss 362.0076 l_rec 175.3620 l_s 622.1522 l_aux 0.0000 l1x 1.00 dead 0/8192
775
+ [epoch 63 step 1600] loss 359.9676 l_rec 173.0841 l_s 622.9447 l_aux 0.0000 l1x 1.00 dead 0/8192
776
+ [epoch 67 step 1700] loss 353.2803 l_rec 169.9748 l_s 611.0184 l_aux 0.0000 l1x 1.00 dead 0/8192
777
+ [epoch 71 step 1800] loss 351.8120 l_rec 168.2344 l_s 611.9254 l_aux 0.0000 l1x 1.00 dead 0/8192
778
+ [epoch 75 step 1900] loss 349.1650 l_rec 166.5981 l_s 608.5565 l_aux 0.0000 l1x 1.00 dead 0/8192
779
+ [epoch 79 step 2000] loss 346.7988 l_rec 164.8113 l_s 606.6248 l_aux 0.0000 l1x 1.00 dead 0/8192
780
+ [epoch 83 step 2100] loss 345.5477 l_rec 164.1883 l_s 604.5311 l_aux 0.0000 l1x 1.00 dead 0/8192
781
+ [epoch 87 step 2200] loss 340.5246 l_rec 161.5341 l_s 596.6351 l_aux 0.0000 l1x 1.00 dead 0/8192
782
+ [epoch 91 step 2300] loss 343.8881 l_rec 162.2247 l_s 605.5447 l_aux 0.0000 l1x 1.00 dead 0/8192
783
+ [epoch 95 step 2400] loss 334.8434 l_rec 159.4394 l_s 584.6799 l_aux 0.0000 l1x 1.00 dead 0/8192
784
+ [epoch 99 step 2500] loss 337.3550 l_rec 159.9777 l_s 591.2577 l_aux 0.0000 l1x 1.00 dead 0/8192
785
+ [step 75000] EV=0.7904 L0=1004.1 (38.6s)
786
+ [step 75000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step75000.pt
787
+ [step 89000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
788
+ [step 89000] preprocess center_scale: mean-shift=0.254 scale=0.0335
789
+ [epoch 3 step 100] loss 748.0341 l_rec 747.6692 l_s 3.0713 l_aux 0.0000 l1x 0.40 dead 0/8192
790
+ [epoch 7 step 200] loss 568.7213 l_rec 513.8453 l_s 229.7988 l_aux 0.0000 l1x 0.80 dead 0/8192
791
+ [epoch 11 step 300] loss 468.1664 l_rec 344.7121 l_s 411.5143 l_aux 0.0000 l1x 1.00 dead 0/8192
792
+ [epoch 15 step 400] loss 415.8055 l_rec 255.8751 l_s 533.1014 l_aux 0.0000 l1x 1.00 dead 0/8192
793
+ [epoch 19 step 500] loss 400.7081 l_rec 224.8140 l_s 586.3135 l_aux 0.0000 l1x 1.00 dead 0/8192
794
+ [epoch 23 step 600] loss 396.2085 l_rec 213.2759 l_s 609.7754 l_aux 0.0000 l1x 1.00 dead 0/8192
795
+ [epoch 27 step 700] loss 395.4067 l_rec 206.1863 l_s 630.7345 l_aux 0.0000 l1x 1.00 dead 0/8192
796
+ [epoch 31 step 800] loss 389.4592 l_rec 200.1281 l_s 631.1038 l_aux 0.0000 l1x 1.00 dead 0/8192
797
+ [epoch 35 step 900] loss 378.3689 l_rec 193.3715 l_s 616.6580 l_aux 0.0000 l1x 1.00 dead 0/8192
798
+ [epoch 39 step 1000] loss 380.5645 l_rec 190.6112 l_s 633.1776 l_aux 0.0000 l1x 1.00 dead 0/8192
799
+ [epoch 43 step 1100] loss 368.3098 l_rec 184.1561 l_s 613.8455 l_aux 0.0000 l1x 1.00 dead 0/8192
800
+ [epoch 47 step 1200] loss 361.1805 l_rec 180.1357 l_s 603.4827 l_aux 0.0000 l1x 1.00 dead 0/8192
801
+ [epoch 51 step 1300] loss 364.2720 l_rec 178.8879 l_s 617.9470 l_aux 0.0000 l1x 1.00 dead 0/8192
802
+ [epoch 55 step 1400] loss 361.8168 l_rec 175.9345 l_s 619.6076 l_aux 0.0000 l1x 1.00 dead 0/8192
803
+ [epoch 59 step 1500] loss 359.5590 l_rec 173.9451 l_s 618.7129 l_aux 0.0000 l1x 1.00 dead 0/8192
804
+ [epoch 63 step 1600] loss 357.7551 l_rec 171.8373 l_s 619.7258 l_aux 0.0000 l1x 1.00 dead 0/8192
805
+ [epoch 67 step 1700] loss 351.4660 l_rec 168.9047 l_s 608.5376 l_aux 0.0000 l1x 1.00 dead 0/8192
806
+ [epoch 71 step 1800] loss 349.3088 l_rec 166.8289 l_s 608.2664 l_aux 0.0000 l1x 1.00 dead 0/8192
807
+ [epoch 75 step 1900] loss 346.9173 l_rec 165.1031 l_s 606.0473 l_aux 0.0000 l1x 1.00 dead 0/8192
808
+ [epoch 79 step 2000] loss 344.6494 l_rec 163.3597 l_s 604.2991 l_aux 0.0000 l1x 1.00 dead 0/8192
809
+ [epoch 83 step 2100] loss 343.5021 l_rec 162.5884 l_s 603.0457 l_aux 0.0000 l1x 1.00 dead 0/8192
810
+ [epoch 87 step 2200] loss 338.1418 l_rec 159.9830 l_s 593.8627 l_aux 0.0000 l1x 1.00 dead 0/8192
811
+ [epoch 91 step 2300] loss 341.4808 l_rec 160.9616 l_s 601.7305 l_aux 0.0000 l1x 1.00 dead 0/8192
812
+ [epoch 95 step 2400] loss 332.4887 l_rec 157.8400 l_s 582.1622 l_aux 0.0000 l1x 1.00 dead 0/8192
813
+ [epoch 99 step 2500] loss 335.1154 l_rec 158.3339 l_s 589.2714 l_aux 0.0000 l1x 1.00 dead 0/8192
814
+ [step 89000] EV=0.7924 L0=1002.0 (38.3s)
815
+ [step 89000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step89000.pt
816
+ [step 102000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
817
+ [step 102000] preprocess center_scale: mean-shift=0.846 scale=0.0337
818
+ [epoch 3 step 100] loss 749.1191 l_rec 748.7728 l_s 2.9146 l_aux 0.0000 l1x 0.40 dead 0/8192
819
+ [epoch 7 step 200] loss 541.7220 l_rec 472.3496 l_s 290.5042 l_aux 0.0000 l1x 0.80 dead 0/8192
820
+ [epoch 11 step 300] loss 462.3163 l_rec 334.9628 l_s 424.5118 l_aux 0.0000 l1x 1.00 dead 0/8192
821
+ [epoch 15 step 400] loss 411.3479 l_rec 262.1570 l_s 497.3030 l_aux 0.0000 l1x 1.00 dead 0/8192
822
+ [epoch 19 step 500] loss 391.6809 l_rec 225.5022 l_s 553.9291 l_aux 0.0000 l1x 1.00 dead 0/8192
823
+ [epoch 23 step 600] loss 384.6104 l_rec 210.3283 l_s 580.9403 l_aux 0.0000 l1x 1.00 dead 0/8192
824
+ [epoch 27 step 700] loss 381.7610 l_rec 201.5428 l_s 600.7274 l_aux 0.0000 l1x 1.00 dead 0/8192
825
+ [epoch 31 step 800] loss 378.0996 l_rec 195.6272 l_s 608.2411 l_aux 0.0000 l1x 1.00 dead 0/8192
826
+ [epoch 35 step 900] loss 365.6752 l_rec 188.0760 l_s 591.9971 l_aux 0.0000 l1x 1.00 dead 0/8192
827
+ [epoch 39 step 1000] loss 369.0549 l_rec 185.4153 l_s 612.1321 l_aux 0.0000 l1x 1.00 dead 0/8192
828
+ [epoch 43 step 1100] loss 356.7156 l_rec 178.6516 l_s 593.5465 l_aux 0.0000 l1x 1.00 dead 0/8192
829
+ [epoch 47 step 1200] loss 349.9706 l_rec 174.5347 l_s 584.7863 l_aux 0.0000 l1x 1.00 dead 0/8192
830
+ [epoch 51 step 1300] loss 351.7260 l_rec 173.4003 l_s 594.4187 l_aux 0.0000 l1x 1.00 dead 0/8192
831
+ [epoch 55 step 1400] loss 351.2224 l_rec 170.7159 l_s 601.6884 l_aux 0.0000 l1x 1.00 dead 0/8192
832
+ [epoch 59 step 1500] loss 346.8577 l_rec 168.0294 l_s 596.0942 l_aux 0.0000 l1x 1.00 dead 0/8192
833
+ [epoch 63 step 1600] loss 344.8054 l_rec 166.1112 l_s 595.6473 l_aux 0.0000 l1x 1.00 dead 0/8192
834
+ [epoch 67 step 1700] loss 340.3484 l_rec 163.2158 l_s 590.4421 l_aux 0.0000 l1x 1.00 dead 0/8192
835
+ [epoch 71 step 1800] loss 338.3710 l_rec 161.3917 l_s 589.9309 l_aux 0.0000 l1x 1.00 dead 0/8192
836
+ [epoch 75 step 1900] loss 336.0672 l_rec 159.7463 l_s 587.7364 l_aux 0.0000 l1x 1.00 dead 0/8192
837
+ [epoch 79 step 2000] loss 332.8061 l_rec 157.8021 l_s 583.3464 l_aux 0.0000 l1x 1.00 dead 0/8192
838
+ [epoch 83 step 2100] loss 331.1163 l_rec 157.0106 l_s 580.3522 l_aux 0.0000 l1x 1.00 dead 0/8192
839
+ [epoch 87 step 2200] loss 326.0347 l_rec 154.3628 l_s 572.2397 l_aux 0.0000 l1x 1.00 dead 0/8192
840
+ [epoch 91 step 2300] loss 329.8582 l_rec 155.1155 l_s 582.4758 l_aux 0.0000 l1x 1.00 dead 0/8192
841
+ [epoch 95 step 2400] loss 321.5541 l_rec 151.8258 l_s 565.7611 l_aux 0.0000 l1x 1.00 dead 0/8192
842
+ [epoch 99 step 2500] loss 323.1918 l_rec 152.7396 l_s 568.1740 l_aux 0.0000 l1x 1.00 dead 0/8192
843
+ [step 102000] EV=0.7998 L0=983.0 (39.0s)
844
+ [step 102000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step102000.pt
845
+ [step 116000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
846
+ [step 116000] preprocess center_scale: mean-shift=1.889 scale=0.0357
847
+ [epoch 3 step 100] loss 724.9061 l_rec 724.0076 l_s 7.5624 l_aux 0.0000 l1x 0.40 dead 0/8192
848
+ [epoch 7 step 200] loss 511.7272 l_rec 431.3931 l_s 336.4077 l_aux 0.0000 l1x 0.80 dead 0/8192
849
+ [epoch 11 step 300] loss 449.1479 l_rec 327.9159 l_s 404.1068 l_aux 0.0000 l1x 1.00 dead 0/8192
850
+ [epoch 15 step 400] loss 403.6541 l_rec 270.6779 l_s 443.2538 l_aux 0.0000 l1x 1.00 dead 0/8192
851
+ [epoch 19 step 500] loss 380.5852 l_rec 234.0477 l_s 488.4583 l_aux 0.0000 l1x 1.00 dead 0/8192
852
+ [epoch 23 step 600] loss 369.0148 l_rec 214.6943 l_s 514.4016 l_aux 0.0000 l1x 1.00 dead 0/8192
853
+ [epoch 27 step 700] loss 361.8515 l_rec 201.5356 l_s 534.3863 l_aux 0.0000 l1x 1.00 dead 0/8192
854
+ [epoch 31 step 800] loss 358.6213 l_rec 193.8765 l_s 549.1495 l_aux 0.0000 l1x 1.00 dead 0/8192
855
+ [epoch 35 step 900] loss 344.6246 l_rec 183.6483 l_s 536.5878 l_aux 0.0000 l1x 1.00 dead 0/8192
856
+ [epoch 39 step 1000] loss 348.1734 l_rec 180.0685 l_s 560.3496 l_aux 0.0000 l1x 1.00 dead 0/8192
857
+ [epoch 43 step 1100] loss 336.4772 l_rec 172.3906 l_s 546.9556 l_aux 0.0000 l1x 1.00 dead 0/8192
858
+ [epoch 47 step 1200] loss 330.1644 l_rec 167.7282 l_s 541.4540 l_aux 0.0000 l1x 1.00 dead 0/8192
859
+ [epoch 51 step 1300] loss 330.5358 l_rec 165.7055 l_s 549.4342 l_aux 0.0000 l1x 1.00 dead 0/8192
860
+ [epoch 55 step 1400] loss 331.1413 l_rec 163.2043 l_s 559.7899 l_aux 0.0000 l1x 1.00 dead 0/8192
861
+ [epoch 59 step 1500] loss 325.1277 l_rec 159.6311 l_s 551.6555 l_aux 0.0000 l1x 1.00 dead 0/8192
862
+ [epoch 63 step 1600] loss 322.3908 l_rec 157.5840 l_s 549.3562 l_aux 0.0000 l1x 1.00 dead 0/8192
863
+ [epoch 67 step 1700] loss 319.7092 l_rec 154.8261 l_s 549.6100 l_aux 0.0000 l1x 1.00 dead 0/8192
864
+ [epoch 71 step 1800] loss 318.0562 l_rec 152.7451 l_s 551.0366 l_aux 0.0000 l1x 1.00 dead 0/8192
865
+ [epoch 75 step 1900] loss 315.6389 l_rec 151.0626 l_s 548.5874 l_aux 0.0000 l1x 1.00 dead 0/8192
866
+ [epoch 79 step 2000] loss 311.6722 l_rec 148.7762 l_s 542.9866 l_aux 0.0000 l1x 1.00 dead 0/8192
867
+ [epoch 83 step 2100] loss 309.5476 l_rec 147.7546 l_s 539.3099 l_aux 0.0000 l1x 1.00 dead 0/8192
868
+ [epoch 87 step 2200] loss 304.7930 l_rec 145.1824 l_s 532.0354 l_aux 0.0000 l1x 1.00 dead 0/8192
869
+ [epoch 91 step 2300] loss 308.0286 l_rec 145.8863 l_s 540.4742 l_aux 0.0000 l1x 1.00 dead 0/8192
870
+ [epoch 95 step 2400] loss 300.9712 l_rec 142.8587 l_s 527.0415 l_aux 0.0000 l1x 1.00 dead 0/8192
871
+ [epoch 99 step 2500] loss 302.1165 l_rec 143.5032 l_s 528.7108 l_aux 0.0000 l1x 1.00 dead 0/8192
872
+ [step 116000] EV=0.8121 L0=957.7 (38.4s)
873
+ [step 116000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step116000.pt
874
+ [step 130000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
875
+ [step 130000] preprocess center_scale: mean-shift=2.523 scale=0.0366
876
+ [epoch 3 step 100] loss 701.0851 l_rec 699.6489 l_s 12.0901 l_aux 0.0000 l1x 0.40 dead 0/8192
877
+ [epoch 7 step 200] loss 502.8510 l_rec 432.0868 l_s 296.3325 l_aux 0.0000 l1x 0.80 dead 0/8192
878
+ [epoch 11 step 300] loss 440.8058 l_rec 329.0974 l_s 372.3613 l_aux 0.0000 l1x 1.00 dead 0/8192
879
+ [epoch 15 step 400] loss 398.5544 l_rec 274.2242 l_s 414.4341 l_aux 0.0000 l1x 1.00 dead 0/8192
880
+ [epoch 19 step 500] loss 376.6007 l_rec 241.3489 l_s 450.8392 l_aux 0.0000 l1x 1.00 dead 0/8192
881
+ [epoch 23 step 600] loss 362.8257 l_rec 222.2321 l_s 468.6452 l_aux 0.0000 l1x 1.00 dead 0/8192
882
+ [epoch 27 step 700] loss 352.9850 l_rec 207.3141 l_s 485.5698 l_aux 0.0000 l1x 1.00 dead 0/8192
883
+ [epoch 31 step 800] loss 348.7554 l_rec 198.1585 l_s 501.9896 l_aux 0.0000 l1x 1.00 dead 0/8192
884
+ [epoch 35 step 900] loss 333.3333 l_rec 185.4000 l_s 493.1107 l_aux 0.0000 l1x 1.00 dead 0/8192
885
+ [epoch 39 step 1000] loss 336.0175 l_rec 180.8955 l_s 517.0736 l_aux 0.0000 l1x 1.00 dead 0/8192
886
+ [epoch 43 step 1100] loss 323.9827 l_rec 171.9637 l_s 506.7300 l_aux 0.0000 l1x 1.00 dead 0/8192
887
+ [epoch 47 step 1200] loss 317.4915 l_rec 166.2360 l_s 504.1848 l_aux 0.0000 l1x 1.00 dead 0/8192
888
+ [epoch 51 step 1300] loss 316.7279 l_rec 162.9697 l_s 512.5274 l_aux 0.0000 l1x 1.00 dead 0/8192
889
+ [epoch 55 step 1400] loss 317.5267 l_rec 160.1908 l_s 524.4530 l_aux 0.0000 l1x 1.00 dead 0/8192
890
+ [epoch 59 step 1500] loss 310.5512 l_rec 155.7658 l_s 515.9512 l_aux 0.0000 l1x 1.00 dead 0/8192
891
+ [epoch 63 step 1600] loss 307.5986 l_rec 153.2803 l_s 514.3945 l_aux 0.0000 l1x 1.00 dead 0/8192
892
+ [epoch 67 step 1700] loss 305.3134 l_rec 150.3488 l_s 516.5486 l_aux 0.0000 l1x 1.00 dead 0/8192
893
+ [epoch 71 step 1800] loss 303.7579 l_rec 147.9983 l_s 519.1989 l_aux 0.0000 l1x 1.00 dead 0/8192
894
+ [epoch 75 step 1900] loss 301.1042 l_rec 146.1700 l_s 516.4471 l_aux 0.0000 l1x 1.00 dead 0/8192
895
+ [epoch 79 step 2000] loss 296.9882 l_rec 143.5553 l_s 511.4429 l_aux 0.0000 l1x 1.00 dead 0/8192
896
+ [epoch 83 step 2100] loss 294.8148 l_rec 142.2902 l_s 508.4154 l_aux 0.0000 l1x 1.00 dead 0/8192
897
+ [epoch 87 step 2200] loss 290.0573 l_rec 139.5711 l_s 501.6205 l_aux 0.0000 l1x 1.00 dead 0/8192
898
+ [epoch 91 step 2300] loss 293.0697 l_rec 140.2375 l_s 509.4405 l_aux 0.0000 l1x 1.00 dead 0/8192
899
+ [epoch 95 step 2400] loss 287.0190 l_rec 137.4729 l_s 498.4872 l_aux 0.0000 l1x 1.00 dead 0/8192
900
+ [epoch 99 step 2500] loss 287.3862 l_rec 137.8345 l_s 498.5059 l_aux 0.0000 l1x 1.00 dead 0/8192
901
+ [step 130000] EV=0.8196 L0=940.0 (38.5s)
902
+ [step 130000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step130000.pt
903
+ [step 143000] loading W_U from /home/mh2274/rds/hpc-work/wu_crosscoder/snapshots
904
+ [step 143000] preprocess center_scale: mean-shift=3.011 scale=0.0368
905
+ [epoch 3 step 100] loss 675.9028 l_rec 673.9046 l_s 16.8199 l_aux 0.0000 l1x 0.40 dead 0/8192
906
+ [epoch 7 step 200] loss 493.9295 l_rec 431.2505 l_s 262.4747 l_aux 0.0000 l1x 0.80 dead 1/8192
907
+ [epoch 11 step 300] loss 432.0815 l_rec 328.5571 l_s 345.0815 l_aux 0.0000 l1x 1.00 dead 1/8192
908
+ [epoch 15 step 400] loss 393.6248 l_rec 277.0956 l_s 388.4308 l_aux 0.0000 l1x 1.00 dead 1/8192
909
+ [epoch 19 step 500] loss 373.4698 l_rec 248.0745 l_s 417.9845 l_aux 0.0000 l1x 1.00 dead 1/8192
910
+ [epoch 23 step 600] loss 358.6503 l_rec 229.4060 l_s 430.8142 l_aux 0.0000 l1x 1.00 dead 1/8192
911
+ [epoch 27 step 700] loss 347.6536 l_rec 213.6505 l_s 446.6770 l_aux 0.0000 l1x 1.00 dead 1/8192
912
+ [epoch 31 step 800] loss 343.1483 l_rec 203.8816 l_s 464.2223 l_aux 0.0000 l1x 1.00 dead 1/8192
913
+ [epoch 35 step 900] loss 326.7617 l_rec 189.4266 l_s 457.7834 l_aux 0.0000 l1x 1.00 dead 1/8192
914
+ [epoch 39 step 1000] loss 328.7961 l_rec 184.3490 l_s 481.4905 l_aux 0.0000 l1x 1.00 dead 1/8192
915
+ [epoch 43 step 1100] loss 316.5259 l_rec 174.4823 l_s 473.4785 l_aux 0.0000 l1x 1.00 dead 1/8192
916
+ [epoch 47 step 1200] loss 309.8365 l_rec 168.0535 l_s 472.6097 l_aux 0.0000 l1x 1.00 dead 1/8192
917
+ [epoch 51 step 1300] loss 307.8077 l_rec 163.8541 l_s 479.8455 l_aux 0.0000 l1x 1.00 dead 1/8192
918
+ [epoch 55 step 1400] loss 308.2553 l_rec 160.7455 l_s 491.6992 l_aux 0.0000 l1x 1.00 dead 1/8192
919
+ [epoch 59 step 1500] loss 300.4465 l_rec 155.2913 l_s 483.8507 l_aux 0.0000 l1x 1.00 dead 1/8192
920
+ [epoch 63 step 1600] loss 296.9078 l_rec 152.1746 l_s 482.4443 l_aux 0.0000 l1x 1.00 dead 1/8192
921
+ [epoch 67 step 1700] loss 294.7872 l_rec 148.8271 l_s 486.5336 l_aux 0.0000 l1x 1.00 dead 1/8192
922
+ [epoch 71 step 1800] loss 293.0072 l_rec 146.0649 l_s 489.8075 l_aux 0.0000 l1x 1.00 dead 1/8192
923
+ [epoch 75 step 1900] loss 289.9734 l_rec 143.6024 l_s 487.9035 l_aux 0.0000 l1x 1.00 dead 0/8192
924
+ [epoch 79 step 2000] loss 285.5460 l_rec 140.6933 l_s 482.8422 l_aux 0.0000 l1x 1.00 dead 0/8192
925
+ [epoch 83 step 2100] loss 283.1811 l_rec 138.9506 l_s 480.7684 l_aux 0.0000 l1x 1.00 dead 0/8192
926
+ [epoch 87 step 2200] loss 278.4196 l_rec 136.1251 l_s 474.3150 l_aux 0.0000 l1x 1.00 dead 0/8192
927
+ [epoch 91 step 2300] loss 281.3252 l_rec 136.6572 l_s 482.2265 l_aux 0.0000 l1x 1.00 dead 0/8192
928
+ [epoch 95 step 2400] loss 275.9033 l_rec 134.0996 l_s 472.6790 l_aux 0.0000 l1x 1.00 dead 0/8192
929
+ [epoch 99 step 2500] loss 275.8053 l_rec 134.1536 l_s 472.1723 l_aux 0.0000 l1x 1.00 dead 0/8192
930
+ [step 143000] EV=0.8245 L0=924.1 (39.0s)
931
+ [step 143000] saved /home/mh2274/rds/hpc-work/unembedding_development/csd3_results/t1_3_per_snap_sae/20260428T230005Z_job28571218/wu_sae_dsae8192_step143000.pt
932
+
933
+ === T1.3 done: 32 snapshots in 21.5 min ===
934
+ step 0: EV=0.7289 L0=1150.1 (40.3s) -> wu_sae_dsae8192_step0.pt
935
+ step 1: EV=0.7289 L0=1150.1 (38.5s) -> wu_sae_dsae8192_step1.pt
936
+ step 2: EV=0.7289 L0=1150.1 (38.4s) -> wu_sae_dsae8192_step2.pt
937
+ step 4: EV=0.7289 L0=1150.1 (38.5s) -> wu_sae_dsae8192_step4.pt
938
+ step 8: EV=0.7289 L0=1150.1 (38.4s) -> wu_sae_dsae8192_step8.pt
939
+ step 16: EV=0.7289 L0=1150.1 (38.9s) -> wu_sae_dsae8192_step16.pt
940
+ step 32: EV=0.7289 L0=1150.1 (38.6s) -> wu_sae_dsae8192_step32.pt
941
+ step 64: EV=0.7289 L0=1150.1 (38.5s) -> wu_sae_dsae8192_step64.pt
942
+ step 128: EV=0.7290 L0=1149.6 (38.4s) -> wu_sae_dsae8192_step128.pt
943
+ step 256: EV=0.7317 L0=1142.4 (38.6s) -> wu_sae_dsae8192_step256.pt
944
+ step 512: EV=0.7543 L0=1086.6 (38.4s) -> wu_sae_dsae8192_step512.pt
945
+ step 1000: EV=0.7862 L0=996.9 (38.8s) -> wu_sae_dsae8192_step1000.pt
946
+ step 2000: EV=0.7993 L0=969.2 (38.4s) -> wu_sae_dsae8192_step2000.pt
947
+ step 3000: EV=0.7993 L0=972.4 (38.5s) -> wu_sae_dsae8192_step3000.pt
948
+ step 4000: EV=0.7978 L0=977.4 (38.6s) -> wu_sae_dsae8192_step4000.pt
949
+ step 5000: EV=0.7961 L0=981.8 (39.4s) -> wu_sae_dsae8192_step5000.pt
950
+ step 6000: EV=0.7948 L0=985.5 (38.6s) -> wu_sae_dsae8192_step6000.pt
951
+ step 7000: EV=0.7938 L0=988.1 (38.4s) -> wu_sae_dsae8192_step7000.pt
952
+ step 8000: EV=0.7931 L0=990.1 (38.7s) -> wu_sae_dsae8192_step8000.pt
953
+ step 9000: EV=0.7924 L0=992.0 (38.3s) -> wu_sae_dsae8192_step9000.pt
954
+ step 14000: EV=0.7912 L0=996.5 (38.5s) -> wu_sae_dsae8192_step14000.pt
955
+ step 21000: EV=0.7908 L0=998.5 (38.4s) -> wu_sae_dsae8192_step21000.pt
956
+ step 27000: EV=0.7908 L0=999.3 (39.2s) -> wu_sae_dsae8192_step27000.pt
957
+ step 34000: EV=0.7906 L0=999.5 (38.5s) -> wu_sae_dsae8192_step34000.pt
958
+ step 47000: EV=0.7906 L0=1000.3 (38.4s) -> wu_sae_dsae8192_step47000.pt
959
+ step 61000: EV=0.7904 L0=1001.8 (39.2s) -> wu_sae_dsae8192_step61000.pt
960
+ step 75000: EV=0.7904 L0=1004.1 (38.6s) -> wu_sae_dsae8192_step75000.pt
961
+ step 89000: EV=0.7924 L0=1002.0 (38.3s) -> wu_sae_dsae8192_step89000.pt
962
+ step 102000: EV=0.7998 L0=983.0 (39.0s) -> wu_sae_dsae8192_step102000.pt
963
+ step 116000: EV=0.8121 L0=957.7 (38.4s) -> wu_sae_dsae8192_step116000.pt
964
+ step 130000: EV=0.8196 L0=940.0 (38.5s) -> wu_sae_dsae8192_step130000.pt
965
+ step 143000: EV=0.8245 L0=924.1 (39.0s) -> wu_sae_dsae8192_step143000.pt
evaluation/eval-corpus/README.md ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ # Evaluation token corpus
2
+
3
+ `eval_tokens.pt` — the multi-script held-out evaluation corpus used by the
4
+ temporal-localization / intervention experiments (`--eval-tokens` input).
5
+ Tokenized (Pythia BPE) short slices per language, built by streaming
6
+ `wikimedia/wikipedia` (CC-BY-SA 4.0; text attribution: Wikipedia contributors).
7
+ The builder script ships with the code release
8
+ (`experiments/causal/temporal_localization_patching/`; see `docs/DATA.md`).
evaluation/eval-corpus/eval_tokens.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7ddc3522155902938eb25202a1c677cb413906b63af24605313edbf5ce38a7a3
3
+ size 2091077
olmo-2-7b/W_U/cross-snapshot-32/d32768/seed0.config.json ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "public_label": "OLMo-2-7B W_U crosscoder, 32 snapshots, d_sae=32768, seed 0",
3
+ "config": {
4
+ "sae_type": "crosscoder",
5
+ "d_model": 4096,
6
+ "expansion_factor": 8.0,
7
+ "use_decoder_bias": true,
8
+ "act_fn": "jumprelu",
9
+ "norm_activation": "inference",
10
+ "sparsity_include_decoder_norm": true,
11
+ "top_k": 50,
12
+ "use_triton_kernel": false,
13
+ "sparsity_threshold_for_triton_spmm_kernel": 0.996,
14
+ "jumprelu_threshold_window": 4.0,
15
+ "hook_points": [
16
+ "wu_step_0",
17
+ "wu_step_1",
18
+ "wu_step_2",
19
+ "wu_step_3",
20
+ "wu_step_4",
21
+ "wu_step_5",
22
+ "wu_step_6",
23
+ "wu_step_7",
24
+ "wu_step_8",
25
+ "wu_step_9",
26
+ "wu_step_10",
27
+ "wu_step_11",
28
+ "wu_step_12",
29
+ "wu_step_13",
30
+ "wu_step_14",
31
+ "wu_step_15",
32
+ "wu_step_16",
33
+ "wu_step_17",
34
+ "wu_step_18",
35
+ "wu_step_19",
36
+ "wu_step_20",
37
+ "wu_step_21",
38
+ "wu_step_22",
39
+ "wu_step_23",
40
+ "wu_step_24",
41
+ "wu_step_25",
42
+ "wu_step_26",
43
+ "wu_step_27",
44
+ "wu_step_28",
45
+ "wu_step_29",
46
+ "wu_step_30",
47
+ "wu_step_31"
48
+ ]
49
+ },
50
+ "training": {
51
+ "lr": 1e-05,
52
+ "l1_coefficient": 0.3,
53
+ "tanh_stretch_coefficient": 1.0,
54
+ "jumprelu_lr_factor": 0.3,
55
+ "n_epochs": 300,
56
+ "batch_size": 1024,
57
+ "expansion_factor": 8.0,
58
+ "d_sae": 32768,
59
+ "input_preprocess": "center_scale",
60
+ "l1_warmup_fraction": 0.1,
61
+ "lr_warmup_fraction": 0.1,
62
+ "lr_decay_fraction": 0.2,
63
+ "init_threshold": 0.1,
64
+ "decoder_transpose_init": 1.0,
65
+ "auxk_coefficient": 0.0,
66
+ "auxk_k": 512,
67
+ "dead_window_steps": 2000,
68
+ "amp_dtype": "fp32",
69
+ "optimizer": "adam"
70
+ },
71
+ "quality": {
72
+ "explained_variance": 0.8526035831389717,
73
+ "reconstruction_mse": 0.1473964168498223,
74
+ "mean_l0": 557.2795322963169,
75
+ "dead_rate": 0.0,
76
+ "d_sae": 32768,
77
+ "n_snapshots": 32,
78
+ "n_rows": 100352
79
+ },
80
+ "steps": [
81
+ 150,
82
+ 600,
83
+ 700,
84
+ 850,
85
+ 900,
86
+ 1000,
87
+ 2000,
88
+ 3000,
89
+ 4000,
90
+ 5000,
91
+ 6000,
92
+ 7000,
93
+ 8000,
94
+ 9000,
95
+ 14000,
96
+ 21000,
97
+ 27000,
98
+ 34000,
99
+ 47000,
100
+ 110000,
101
+ 173000,
102
+ 236000,
103
+ 299000,
104
+ 362000,
105
+ 425000,
106
+ 488000,
107
+ 614000,
108
+ 677000,
109
+ 740000,
110
+ 803000,
111
+ 866000,
112
+ 928000
113
+ ],
114
+ "model_name": "allenai/OLMo-2-1124-7B",
115
+ "seed": 0,
116
+ "kind": "cross-snapshot",
117
+ "preprocess_mode": "center_scale"
118
+ }
olmo-2-7b/W_U/cross-snapshot-32/d32768/seed0.md ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # OLMo-2-7B W_U crosscoder, 32 snapshots, d_sae=32768, seed 0
2
+
3
+ Cross-snapshot W_U crosscoder trained over 32 snapshots of
4
+ allenai/OLMo-2-1124-7B's unembedding matrix. Converted to the unified
5
+ `parameter-trajectory-crosscoders/` safetensors layout from the original
6
+ training-run `.pt` (cc_olmo27b_dsae32768_seed0.pt).
7
+
8
+ Cross-family replication of the Pythia W_U trajectory-crosscoder analysis on allenai/OLMo-2-1124-7B checkpoints.
9
+
10
+ Quality in the sidecar is recomputed post-conversion with the release
11
+ protocol (scripts/eval/recompute_metrics.py); the original training log
12
+ reported EV under a different (train-time) evaluation protocol.
pythia-160m/W_U/lambda-sweep/d8192/lam0p40_seed0.metrics_recomputed.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "recomputed": {
3
+ "explained_variance": 0.7479858934016762,
4
+ "reconstruction_mse": 0.25193381041288376,
5
+ "mean_l0": 160.0767987060547,
6
+ "dead_rate": 0.0,
7
+ "d_sae": 8192,
8
+ "n_snapshots": 32,
9
+ "n_rows": 50304,
10
+ "n_batches": 50,
11
+ "preprocess_mode": "center_scale",
12
+ "model_name": "EleutherAI/pythia-160m",
13
+ "runtime_s": 39.5
14
+ },
15
+ "stored": {
16
+ "explained_variance": 0.7479858952216084,
17
+ "reconstruction_mse": 0.25193381249904634,
18
+ "mean_l0": 160.0768243408203,
19
+ "dead_rate": 0.0,
20
+ "d_sae": 8192,
21
+ "n_snapshots": 32,
22
+ "n_rows": 50304
23
+ }
24
+ }
pythia-160m/W_U/lambda-sweep/d8192/lam1p00_seed0.metrics_recomputed.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "recomputed": {
3
+ "explained_variance": 0.6323049222593329,
4
+ "reconstruction_mse": 0.36757792353630064,
5
+ "mean_l0": 58.498651123046876,
6
+ "dead_rate": 0.0,
7
+ "d_sae": 8192,
8
+ "n_snapshots": 32,
9
+ "n_rows": 50304,
10
+ "n_batches": 50,
11
+ "preprocess_mode": "center_scale",
12
+ "model_name": "EleutherAI/pythia-160m",
13
+ "runtime_s": 44.9
14
+ },
15
+ "stored": {
16
+ "explained_variance": 0.6323049070911215,
17
+ "reconstruction_mse": 0.36757794439792635,
18
+ "mean_l0": 58.49865539550781,
19
+ "dead_rate": 0.0,
20
+ "d_sae": 8192,
21
+ "n_snapshots": 32,
22
+ "n_rows": 50304
23
+ }
24
+ }
pythia-160m/W_U/lambda-sweep/d8192/lam1p20_seed0.metrics_recomputed.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "recomputed": {
3
+ "explained_variance": 0.6047022828542324,
4
+ "reconstruction_mse": 0.3971094989776611,
5
+ "mean_l0": 45.41338134765625,
6
+ "dead_rate": 0.0,
7
+ "d_sae": 8192,
8
+ "n_snapshots": 32,
9
+ "n_rows": 50304,
10
+ "n_batches": 50,
11
+ "preprocess_mode": "center_scale",
12
+ "model_name": "EleutherAI/pythia-160m",
13
+ "runtime_s": 50.8
14
+ },
15
+ "stored": {
16
+ "explained_variance": 0.6027639222075235,
17
+ "reconstruction_mse": 0.39710951745510104,
18
+ "mean_l0": 45.41338317871094,
19
+ "dead_rate": 0.0,
20
+ "d_sae": 8192,
21
+ "n_snapshots": 32,
22
+ "n_rows": 50304
23
+ }
24
+ }
pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed0.metrics_recomputed.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "recomputed": {
3
+ "explained_variance": 0.581972996490729,
4
+ "reconstruction_mse": 0.4178938126564026,
5
+ "mean_l0": 37.97134643554688,
6
+ "dead_rate": 0.0,
7
+ "d_sae": 8192,
8
+ "n_snapshots": 32,
9
+ "n_rows": 50304,
10
+ "n_batches": 50,
11
+ "preprocess_mode": "center_scale",
12
+ "model_name": "EleutherAI/pythia-160m",
13
+ "runtime_s": 51.3
14
+ },
15
+ "stored": {
16
+ "explained_variance": 0.5819729868726617,
17
+ "reconstruction_mse": 0.4178938287496567,
18
+ "mean_l0": 37.971342163085936,
19
+ "dead_rate": 0.0,
20
+ "d_sae": 8192,
21
+ "n_snapshots": 32,
22
+ "n_rows": 50304
23
+ }
24
+ }
pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed1.metrics_recomputed.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "recomputed": {
3
+ "explained_variance": 0.5816520078445061,
4
+ "reconstruction_mse": 0.4182146990299225,
5
+ "mean_l0": 37.7968505859375,
6
+ "dead_rate": 0.0,
7
+ "d_sae": 8192,
8
+ "n_snapshots": 32,
9
+ "n_rows": 50304,
10
+ "n_batches": 50,
11
+ "preprocess_mode": "center_scale",
12
+ "model_name": "EleutherAI/pythia-160m",
13
+ "runtime_s": 57.0
14
+ },
15
+ "stored": {
16
+ "explained_variance": 0.581652004193779,
17
+ "reconstruction_mse": 0.4182147091627121,
18
+ "mean_l0": 37.79685302734375,
19
+ "dead_rate": 0.0,
20
+ "d_sae": 8192,
21
+ "n_snapshots": 32,
22
+ "n_rows": 50304
23
+ }
24
+ }
pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed2.metrics_recomputed.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "recomputed": {
3
+ "explained_variance": 0.5823660355383331,
4
+ "reconstruction_mse": 0.4175008988380432,
5
+ "mean_l0": 37.98096923828125,
6
+ "dead_rate": 0.0,
7
+ "d_sae": 8192,
8
+ "n_snapshots": 32,
9
+ "n_rows": 50304,
10
+ "n_batches": 50,
11
+ "preprocess_mode": "center_scale",
12
+ "model_name": "EleutherAI/pythia-160m",
13
+ "runtime_s": 56.6
14
+ },
15
+ "stored": {
16
+ "explained_variance": 0.582366028895355,
17
+ "reconstruction_mse": 0.4175009119510651,
18
+ "mean_l0": 37.980972900390626,
19
+ "dead_rate": 0.0,
20
+ "d_sae": 8192,
21
+ "n_snapshots": 32,
22
+ "n_rows": 50304
23
+ }
24
+ }
pythia-160m/W_U/lambda-sweep/d8192/lam1p80_seed0.metrics_recomputed.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "recomputed": {
3
+ "explained_variance": 0.527703739457926,
4
+ "reconstruction_mse": 0.47214577853679657,
5
+ "mean_l0": 23.03343505859375,
6
+ "dead_rate": 0.0,
7
+ "d_sae": 8192,
8
+ "n_snapshots": 32,
9
+ "n_rows": 50304,
10
+ "n_batches": 50,
11
+ "preprocess_mode": "center_scale",
12
+ "model_name": "EleutherAI/pythia-160m",
13
+ "runtime_s": 43.9
14
+ },
15
+ "stored": {
16
+ "explained_variance": 0.527703732469858,
17
+ "reconstruction_mse": 0.4721457928419113,
18
+ "mean_l0": 23.033436279296875,
19
+ "dead_rate": 0.0,
20
+ "d_sae": 8192,
21
+ "n_snapshots": 32,
22
+ "n_rows": 50304
23
+ }
24
+ }
pythia-1b/W_U/cross-snapshot-32-matched-window/d24576/seed0.config.json ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "public_label": "Pythia-1B W_U crosscoder, OLMo-matched late-start 32-snapshot window, d_sae=24576, seed 0",
3
+ "config": {
4
+ "sae_type": "crosscoder",
5
+ "d_model": 2048,
6
+ "expansion_factor": 12.0,
7
+ "use_decoder_bias": true,
8
+ "act_fn": "jumprelu",
9
+ "norm_activation": "inference",
10
+ "sparsity_include_decoder_norm": true,
11
+ "top_k": 50,
12
+ "use_triton_kernel": false,
13
+ "sparsity_threshold_for_triton_spmm_kernel": 0.996,
14
+ "jumprelu_threshold_window": 4.0,
15
+ "hook_points": [
16
+ "wu_step_0",
17
+ "wu_step_1",
18
+ "wu_step_2",
19
+ "wu_step_3",
20
+ "wu_step_4",
21
+ "wu_step_5",
22
+ "wu_step_6",
23
+ "wu_step_7",
24
+ "wu_step_8",
25
+ "wu_step_9",
26
+ "wu_step_10",
27
+ "wu_step_11",
28
+ "wu_step_12",
29
+ "wu_step_13",
30
+ "wu_step_14",
31
+ "wu_step_15",
32
+ "wu_step_16",
33
+ "wu_step_17",
34
+ "wu_step_18",
35
+ "wu_step_19",
36
+ "wu_step_20",
37
+ "wu_step_21",
38
+ "wu_step_22",
39
+ "wu_step_23",
40
+ "wu_step_24",
41
+ "wu_step_25",
42
+ "wu_step_26",
43
+ "wu_step_27",
44
+ "wu_step_28",
45
+ "wu_step_29",
46
+ "wu_step_30",
47
+ "wu_step_31"
48
+ ]
49
+ },
50
+ "training": {
51
+ "lr": 5e-05,
52
+ "l1_coefficient": 0.3,
53
+ "tanh_stretch_coefficient": 1.0,
54
+ "jumprelu_lr_factor": 0.1,
55
+ "n_epochs": 300,
56
+ "batch_size": 1024,
57
+ "expansion_factor": 12.0,
58
+ "d_sae": 24576,
59
+ "input_preprocess": "center_scale",
60
+ "l1_warmup_fraction": 0.1,
61
+ "lr_warmup_fraction": 0.1,
62
+ "lr_decay_fraction": 0.2,
63
+ "init_threshold": 0.1,
64
+ "decoder_transpose_init": 1.0,
65
+ "auxk_coefficient": 0.0,
66
+ "auxk_k": 512,
67
+ "dead_window_steps": 2000,
68
+ "amp_dtype": "bf16",
69
+ "optimizer": "adam"
70
+ },
71
+ "quality": {
72
+ "explained_variance": 0.8844067825487556,
73
+ "reconstruction_mse": 0.11559321757396518,
74
+ "mean_l0": 263.98009787989025,
75
+ "dead_rate": 0.0,
76
+ "d_sae": 24576,
77
+ "n_snapshots": 32,
78
+ "n_rows": 50304
79
+ },
80
+ "steps": [
81
+ 256,
82
+ 512,
83
+ 1000,
84
+ 2000,
85
+ 3000,
86
+ 4000,
87
+ 5000,
88
+ 6000,
89
+ 7000,
90
+ 8000,
91
+ 9000,
92
+ 10000,
93
+ 13000,
94
+ 14000,
95
+ 21000,
96
+ 23000,
97
+ 27000,
98
+ 33000,
99
+ 34000,
100
+ 43000,
101
+ 47000,
102
+ 53000,
103
+ 61000,
104
+ 73000,
105
+ 83000,
106
+ 89000,
107
+ 93000,
108
+ 102000,
109
+ 113000,
110
+ 123000,
111
+ 130000,
112
+ 143000
113
+ ],
114
+ "model_name": "EleutherAI/pythia-1b",
115
+ "seed": 0,
116
+ "kind": "cross-snapshot",
117
+ "preprocess_mode": "center_scale"
118
+ }
pythia-1b/W_U/cross-snapshot-32-matched-window/d24576/seed0.md ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ # Pythia-1B W_U crosscoder, OLMo-matched late-start 32-snapshot window, d_sae=24576, seed 0
2
+
3
+ Cross-snapshot W_U crosscoder trained over 32 snapshots of
4
+ EleutherAI/pythia-1b's unembedding matrix. Converted to the unified
5
+ `parameter-trajectory-crosscoders/` safetensors layout from the original
6
+ training-run `.pt` (pythia1b_wu_late_start32_dsae24576_seed0.pt).
7
+
8
+ Control instrument trained on a checkpoint window matched to the OLMo-2-7B schedule (appendix lifecycle diagnostics).
pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0-sparse.config.json ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "public_label": "Pythia-6.9B W_U crosscoder, 32 snapshots, d_sae=32768, seed 0, selected sparse run (lambda=0.6)",
3
+ "config": {
4
+ "sae_type": "crosscoder",
5
+ "d_model": 4096,
6
+ "expansion_factor": 8.0,
7
+ "use_decoder_bias": true,
8
+ "act_fn": "jumprelu",
9
+ "norm_activation": "inference",
10
+ "sparsity_include_decoder_norm": true,
11
+ "top_k": 50,
12
+ "use_triton_kernel": false,
13
+ "sparsity_threshold_for_triton_spmm_kernel": 0.996,
14
+ "jumprelu_threshold_window": 4.0,
15
+ "hook_points": [
16
+ "wu_step_0",
17
+ "wu_step_1",
18
+ "wu_step_2",
19
+ "wu_step_3",
20
+ "wu_step_4",
21
+ "wu_step_5",
22
+ "wu_step_6",
23
+ "wu_step_7",
24
+ "wu_step_8",
25
+ "wu_step_9",
26
+ "wu_step_10",
27
+ "wu_step_11",
28
+ "wu_step_12",
29
+ "wu_step_13",
30
+ "wu_step_14",
31
+ "wu_step_15",
32
+ "wu_step_16",
33
+ "wu_step_17",
34
+ "wu_step_18",
35
+ "wu_step_19",
36
+ "wu_step_20",
37
+ "wu_step_21",
38
+ "wu_step_22",
39
+ "wu_step_23",
40
+ "wu_step_24",
41
+ "wu_step_25",
42
+ "wu_step_26",
43
+ "wu_step_27",
44
+ "wu_step_28",
45
+ "wu_step_29",
46
+ "wu_step_30",
47
+ "wu_step_31"
48
+ ]
49
+ },
50
+ "training": {
51
+ "lr": 1e-05,
52
+ "l1_coefficient": 0.6,
53
+ "tanh_stretch_coefficient": 1.0,
54
+ "jumprelu_lr_factor": 0.4,
55
+ "n_epochs": 300,
56
+ "batch_size": 1024,
57
+ "expansion_factor": 8.0,
58
+ "d_sae": 32768,
59
+ "input_preprocess": "center_scale",
60
+ "l1_warmup_fraction": 0.05,
61
+ "lr_warmup_fraction": 0.0,
62
+ "lr_decay_fraction": 0.0,
63
+ "init_threshold": 0.15,
64
+ "decoder_transpose_init": 1.0,
65
+ "auxk_coefficient": 0.0,
66
+ "auxk_k": 512,
67
+ "dead_window_steps": 2000,
68
+ "amp_dtype": "fp32",
69
+ "optimizer": "adam"
70
+ },
71
+ "quality": {
72
+ "explained_variance": 0.8080785315044828,
73
+ "reconstruction_mse": 0.19192146819065972,
74
+ "mean_l0": 741.8141662287833,
75
+ "dead_rate": 0.0,
76
+ "d_sae": 32768,
77
+ "n_snapshots": 32,
78
+ "n_rows": 50432
79
+ },
80
+ "steps": [
81
+ 0,
82
+ 1,
83
+ 2,
84
+ 4,
85
+ 8,
86
+ 16,
87
+ 32,
88
+ 64,
89
+ 128,
90
+ 256,
91
+ 512,
92
+ 1000,
93
+ 2000,
94
+ 3000,
95
+ 4000,
96
+ 5000,
97
+ 6000,
98
+ 7000,
99
+ 8000,
100
+ 9000,
101
+ 14000,
102
+ 21000,
103
+ 27000,
104
+ 34000,
105
+ 47000,
106
+ 61000,
107
+ 75000,
108
+ 89000,
109
+ 102000,
110
+ 116000,
111
+ 130000,
112
+ 143000
113
+ ],
114
+ "model_name": "EleutherAI/pythia-6.9b",
115
+ "seed": 0,
116
+ "kind": "cross-snapshot",
117
+ "preprocess_mode": "center_scale"
118
+ }
pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0-sparse.md ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ # Pythia-6.9B W_U crosscoder, 32 snapshots, d_sae=32768, seed 0, selected sparse run (lambda=0.6)
2
+
3
+ Cross-snapshot W_U crosscoder trained over 32 snapshots of
4
+ EleutherAI/pythia-6.9b's unembedding matrix. Converted to the unified
5
+ `parameter-trajectory-crosscoders/` safetensors layout from the original
6
+ training-run `.pt` (wu_cc_p69b_dsae32768_seed0.pt).
7
+
8
+ The paper's selected 6.9B instrument (higher sparsity penalty lambda=0.6); the default-lambda comparison run is seed0.safetensors in this directory.
pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0.config.json ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "public_label": "Pythia-6.9B W_U crosscoder, 32 snapshots, d_sae=32768, seed 0",
3
+ "config": {
4
+ "sae_type": "crosscoder",
5
+ "d_model": 4096,
6
+ "expansion_factor": 8.0,
7
+ "use_decoder_bias": true,
8
+ "act_fn": "jumprelu",
9
+ "norm_activation": "inference",
10
+ "sparsity_include_decoder_norm": true,
11
+ "top_k": 50,
12
+ "use_triton_kernel": false,
13
+ "sparsity_threshold_for_triton_spmm_kernel": 0.996,
14
+ "jumprelu_threshold_window": 4.0,
15
+ "hook_points": [
16
+ "wu_step_0",
17
+ "wu_step_1",
18
+ "wu_step_2",
19
+ "wu_step_3",
20
+ "wu_step_4",
21
+ "wu_step_5",
22
+ "wu_step_6",
23
+ "wu_step_7",
24
+ "wu_step_8",
25
+ "wu_step_9",
26
+ "wu_step_10",
27
+ "wu_step_11",
28
+ "wu_step_12",
29
+ "wu_step_13",
30
+ "wu_step_14",
31
+ "wu_step_15",
32
+ "wu_step_16",
33
+ "wu_step_17",
34
+ "wu_step_18",
35
+ "wu_step_19",
36
+ "wu_step_20",
37
+ "wu_step_21",
38
+ "wu_step_22",
39
+ "wu_step_23",
40
+ "wu_step_24",
41
+ "wu_step_25",
42
+ "wu_step_26",
43
+ "wu_step_27",
44
+ "wu_step_28",
45
+ "wu_step_29",
46
+ "wu_step_30",
47
+ "wu_step_31"
48
+ ]
49
+ },
50
+ "training": {
51
+ "lr": 1e-05,
52
+ "l1_coefficient": 0.3,
53
+ "tanh_stretch_coefficient": 1.0,
54
+ "jumprelu_lr_factor": 0.3,
55
+ "n_epochs": 300,
56
+ "batch_size": 1024,
57
+ "expansion_factor": 8.0,
58
+ "d_sae": 32768,
59
+ "input_preprocess": "center_scale",
60
+ "l1_warmup_fraction": 0.1,
61
+ "lr_warmup_fraction": 0.1,
62
+ "lr_decay_fraction": 0.2,
63
+ "init_threshold": 0.1,
64
+ "decoder_transpose_init": 1.0,
65
+ "auxk_coefficient": 0.0,
66
+ "auxk_k": 512,
67
+ "dead_window_steps": 2000,
68
+ "amp_dtype": "fp32",
69
+ "optimizer": "adam"
70
+ },
71
+ "quality": {
72
+ "explained_variance": 0.8331843988901371,
73
+ "reconstruction_mse": 0.1668156008448849,
74
+ "mean_l0": 1956.59887943171,
75
+ "dead_rate": 0.0,
76
+ "d_sae": 32768,
77
+ "n_snapshots": 32,
78
+ "n_rows": 50432
79
+ },
80
+ "steps": [
81
+ 0,
82
+ 1,
83
+ 2,
84
+ 4,
85
+ 8,
86
+ 16,
87
+ 32,
88
+ 64,
89
+ 128,
90
+ 256,
91
+ 512,
92
+ 1000,
93
+ 2000,
94
+ 3000,
95
+ 4000,
96
+ 5000,
97
+ 6000,
98
+ 7000,
99
+ 8000,
100
+ 9000,
101
+ 14000,
102
+ 21000,
103
+ 27000,
104
+ 34000,
105
+ 47000,
106
+ 61000,
107
+ 75000,
108
+ 89000,
109
+ 102000,
110
+ 116000,
111
+ 130000,
112
+ 143000
113
+ ],
114
+ "model_name": "EleutherAI/pythia-6.9b",
115
+ "seed": 0,
116
+ "kind": "wu-crosscoder",
117
+ "preprocess_mode": "center_scale"
118
+ }
pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0.md ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ # Pythia-6.9B W_U crosscoder · d_sae=32768 · seed 0
2
+
3
+ Cross-snapshot W_U crosscoder trained over 32 snapshots of Pythia-6.9B's
4
+ unembedding matrix. Re-saved from the original T2.1 .pt format
5
+ (`wu-crosscoder-trained/t2_1/v6_20260501T233145Z/wu_cc_p69b_dsae32768_seed0.pt`)
6
+ into the unified `parameter-trajectory-crosscoders/` safetensors layout.