Add files using upload-large-folder tool

_sweep_logs/master.log

@@ -0,0 +1,12 @@
+===== topk_k30_s1 (k=30) =====
+===== topk_k30_s2 (k=30) =====
+===== topk_k35_s2 (k=35) =====
+===== topk_k40_s1 (k=40) =====
+===== topk_k40_s2 (k=40) =====
+===== topk_k45_s1 (k=45) =====
+===== topk_k45_s2 (k=45) =====
+===== topk_k48_s2 (k=48) =====
+===== topk_k50_s1 (k=50) =====
+===== topk_k50_s2 (k=50) =====
+===== EXTRA topk_k35_s1 (k=35) mc =====
+ALL_BARRIER_MC_DONE

_sweep_logs/nohup.out

@@ -0,0 +1,90 @@
+===== topk_k30_s1 (k=30) =====
+
+wrote runs/topk_k30_s1/barrier.json
+
+summary: MSE(α=0) = 0.05222  ->  MSE(α=1) = 0.13120
+         max over path = 0.13120  argmax α = 1.000
+
+summary: peak=0.13200  argmax α=1.000
+         endpoints: 0.05251, 0.13200
+===== topk_k30_s2 (k=30) =====
+
+wrote runs/topk_k30_s2/barrier.json
+
+summary: MSE(α=0) = 0.05386  ->  MSE(α=1) = 0.13341
+         max over path = 0.13341  argmax α = 1.000
+
+summary: peak=0.13334  argmax α=1.000
+         endpoints: 0.05225, 0.13334
+===== topk_k35_s2 (k=35) =====
+
+wrote runs/topk_k35_s2/barrier.json
+
+summary: MSE(α=0) = 0.04577  ->  MSE(α=1) = 0.07856
+         max over path = 0.07856  argmax α = 1.000
+
+summary: peak=0.07992  argmax α=1.000
+         endpoints: 0.04629, 0.07992
+===== topk_k40_s1 (k=40) =====
+
+wrote runs/topk_k40_s1/barrier.json
+
+summary: MSE(α=0) = 0.03803  ->  MSE(α=1) = 0.04402
+         max over path = 0.05656  argmax α = 0.450
+
+summary: peak=0.05201  argmax α=0.400
+         endpoints: 0.03904, 0.04542
+===== topk_k40_s2 (k=40) =====
+
+wrote runs/topk_k40_s2/barrier.json
+
+summary: MSE(α=0) = 0.03702  ->  MSE(α=1) = 0.04158
+         max over path = 0.05404  argmax α = 0.450
+
+summary: peak=0.05049  argmax α=0.400
+         endpoints: 0.03750, 0.04304
+===== topk_k45_s1 (k=45) =====
+
+wrote runs/topk_k45_s1/barrier.json
+
+summary: MSE(α=0) = 0.03002  ->  MSE(α=1) = 0.02165
+         max over path = 0.06657  argmax α = 0.500
+
+summary: peak=0.04251  argmax α=0.400
+         endpoints: 0.02965, 0.02010
+===== topk_k45_s2 (k=45) =====
+
+wrote runs/topk_k45_s2/barrier.json
+
+summary: MSE(α=0) = 0.03035  ->  MSE(α=1) = 0.02334
+         max over path = 0.06798  argmax α = 0.500
+
+summary: peak=0.04420  argmax α=0.450
+         endpoints: 0.02988, 0.02408
+===== topk_k48_s2 (k=48) =====
+
+summary: peak=0.03999  argmax α=0.400
+         endpoints: 0.02708, 0.01338
+===== topk_k50_s1 (k=50) =====
+
+wrote runs/topk_k50_s1/barrier.json
+
+summary: MSE(α=0) = 0.02412  ->  MSE(α=1) = 0.00932
+         max over path = 0.05774  argmax α = 0.500
+
+summary: peak=0.03846  argmax α=0.450
+         endpoints: 0.02413, 0.01036
+===== topk_k50_s2 (k=50) =====
+
+wrote runs/topk_k50_s2/barrier.json
+
+summary: MSE(α=0) = 0.02492  ->  MSE(α=1) = 0.01075
+         max over path = 0.06035  argmax α = 0.450
+
+summary: peak=0.03823  argmax α=0.350
+         endpoints: 0.02410, 0.00993
+===== EXTRA topk_k35_s1 (k=35) mc =====
+
+summary: peak=0.07593  argmax α=1.000
+         endpoints: 0.04496, 0.07593
+ALL_BARRIER_MC_DONE

_sweep_logs/topk_k30_s1_barrier.log

@@ -0,0 +1,30 @@
+
+barrier eval: run=runs/topk_k30_s1  k=30  n=5000  alphas=21
+
+Hungarian alignment: mean cos = 0.6010 (0.6010 per scipy)
+  α=0.000  MSE=0.05222
+  α=0.050  MSE=0.05257
+  α=0.100  MSE=0.05410
+  α=0.150  MSE=0.05670
+  α=0.200  MSE=0.06035
+  α=0.250  MSE=0.06481
+  α=0.300  MSE=0.06998
+  α=0.350  MSE=0.07554
+  α=0.400  MSE=0.08108
+  α=0.450  MSE=0.08637
+  α=0.500  MSE=0.09153
+  α=0.550  MSE=0.09655
+  α=0.600  MSE=0.10156
+  α=0.650  MSE=0.10641
+  α=0.700  MSE=0.11103
+  α=0.750  MSE=0.11534
+  α=0.800  MSE=0.11899
+  α=0.850  MSE=0.12208
+  α=0.900  MSE=0.12484
+  α=0.950  MSE=0.12781
+  α=1.000  MSE=0.13120
+
+wrote runs/topk_k30_s1/barrier.json
+
+summary: MSE(α=0) = 0.05222  ->  MSE(α=1) = 0.13120
+         max over path = 0.13120  argmax α = 1.000

_sweep_logs/topk_k30_s2_barrier.log

@@ -0,0 +1,30 @@
+
+barrier eval: run=runs/topk_k30_s2  k=30  n=5000  alphas=21
+
+Hungarian alignment: mean cos = 0.5997 (0.5997 per scipy)
+  α=0.000  MSE=0.05386
+  α=0.050  MSE=0.05409
+  α=0.100  MSE=0.05553
+  α=0.150  MSE=0.05806
+  α=0.200  MSE=0.06159
+  α=0.250  MSE=0.06596
+  α=0.300  MSE=0.07100
+  α=0.350  MSE=0.07646
+  α=0.400  MSE=0.08206
+  α=0.450  MSE=0.08748
+  α=0.500  MSE=0.09287
+  α=0.550  MSE=0.09813
+  α=0.600  MSE=0.10334
+  α=0.650  MSE=0.10837
+  α=0.700  MSE=0.11316
+  α=0.750  MSE=0.11754
+  α=0.800  MSE=0.12121
+  α=0.850  MSE=0.12434
+  α=0.900  MSE=0.12713
+  α=0.950  MSE=0.13001
+  α=1.000  MSE=0.13341
+
+wrote runs/topk_k30_s2/barrier.json
+
+summary: MSE(α=0) = 0.05386  ->  MSE(α=1) = 0.13341
+         max over path = 0.13341  argmax α = 1.000

_sweep_logs/topk_k40_s1_mc.log

@@ -0,0 +1,43 @@
+
+mode connectivity: run=runs/topk_k40_s1  k=40  train α=[0.2, 0.4, 0.5, 0.6, 0.8]
+
+optimizing midpoint dictionary (80 steps, batch 128)...
+  step   1  meanMSE=0.05600   [0.20:0.0436 0.40:0.0578 0.50:0.0611 0.60:0.0611 0.80:0.0564]
+  step  10  meanMSE=0.05563   [0.20:0.0486 0.40:0.0597 0.50:0.0603 0.60:0.0581 0.80:0.0515]
+  step  20  meanMSE=0.04781   [0.20:0.0420 0.40:0.0526 0.50:0.0525 0.60:0.0498 0.80:0.0423]
+  step  30  meanMSE=0.05105   [0.20:0.0463 0.40:0.0557 0.50:0.0553 0.60:0.0523 0.80:0.0456]
+  step  40  meanMSE=0.05143   [0.20:0.0469 0.40:0.0551 0.50:0.0550 0.60:0.0528 0.80:0.0473]
+  step  50  meanMSE=0.04723   [0.20:0.0418 0.40:0.0516 0.50:0.0517 0.60:0.0494 0.80:0.0417]
+  step  60  meanMSE=0.05595   [0.20:0.0500 0.40:0.0590 0.50:0.0592 0.60:0.0571 0.80:0.0545]
+  step  70  meanMSE=0.04839   [0.20:0.0472 0.40:0.0533 0.50:0.0521 0.60:0.0487 0.80:0.0407]
+  step  80  meanMSE=0.04619   [0.20:0.0445 0.40:0.0507 0.50:0.0499 0.60:0.0465 0.80:0.0395]
+
+optimised in 13.0 sec
+
+evaluating optimised curve (21 αs × 5000 samples)
+  α=0.000  MSE=0.03904
+  α=0.050  MSE=0.03948
+  α=0.100  MSE=0.04142
+  α=0.150  MSE=0.04389
+  α=0.200  MSE=0.04638
+  α=0.250  MSE=0.04857
+  α=0.300  MSE=0.05028
+  α=0.350  MSE=0.05146
+  α=0.400  MSE=0.05201
+  α=0.450  MSE=0.05191
+  α=0.500  MSE=0.05122
+  α=0.550  MSE=0.04995
+  α=0.600  MSE=0.04826
+  α=0.650  MSE=0.04631
+  α=0.700  MSE=0.04447
+  α=0.750  MSE=0.04308
+  α=0.800  MSE=0.04241
+  α=0.850  MSE=0.04253
+  α=0.900  MSE=0.04318
+  α=0.950  MSE=0.04408
+  α=1.000  MSE=0.04542
+
+wrote runs/topk_k40_s1/mode_connectivity.json
+
+summary: peak=0.05201  argmax α=0.400
+         endpoints: 0.03904, 0.04542

_sweep_logs/topk_k40_s2_barrier.log

@@ -0,0 +1,30 @@
+
+barrier eval: run=runs/topk_k40_s2  k=40  n=5000  alphas=21
+
+Hungarian alignment: mean cos = 0.6576 (0.6576 per scipy)
+  α=0.000  MSE=0.03702
+  α=0.050  MSE=0.03730
+  α=0.100  MSE=0.03833
+  α=0.150  MSE=0.04001
+  α=0.200  MSE=0.04229
+  α=0.250  MSE=0.04503
+  α=0.300  MSE=0.04802
+  α=0.350  MSE=0.05084
+  α=0.400  MSE=0.05300
+  α=0.450  MSE=0.05404
+  α=0.500  MSE=0.05397
+  α=0.550  MSE=0.05314
+  α=0.600  MSE=0.05162
+  α=0.650  MSE=0.04985
+  α=0.700  MSE=0.04784
+  α=0.750  MSE=0.04588
+  α=0.800  MSE=0.04399
+  α=0.850  MSE=0.04239
+  α=0.900  MSE=0.04129
+  α=0.950  MSE=0.04095
+  α=1.000  MSE=0.04158
+
+wrote runs/topk_k40_s2/barrier.json
+
+summary: MSE(α=0) = 0.03702  ->  MSE(α=1) = 0.04158
+         max over path = 0.05404  argmax α = 0.450

_sweep_logs/topk_k45_s1_barrier.log

@@ -0,0 +1,30 @@
+
+barrier eval: run=runs/topk_k45_s1  k=45  n=5000  alphas=21
+
+Hungarian alignment: mean cos = 0.6868 (0.6868 per scipy)
+  α=0.000  MSE=0.03002
+  α=0.050  MSE=0.03061
+  α=0.100  MSE=0.03255
+  α=0.150  MSE=0.03566
+  α=0.200  MSE=0.03990
+  α=0.250  MSE=0.04514
+  α=0.300  MSE=0.05113
+  α=0.350  MSE=0.05728
+  α=0.400  MSE=0.06258
+  α=0.450  MSE=0.06588
+  α=0.500  MSE=0.06657
+  α=0.550  MSE=0.06490
+  α=0.600  MSE=0.06119
+  α=0.650  MSE=0.05584
+  α=0.700  MSE=0.04892
+  α=0.750  MSE=0.04164
+  α=0.800  MSE=0.03480
+  α=0.850  MSE=0.02899
+  α=0.900  MSE=0.02470
+  α=0.950  MSE=0.02214
+  α=1.000  MSE=0.02165
+
+wrote runs/topk_k45_s1/barrier.json
+
+summary: MSE(α=0) = 0.03002  ->  MSE(α=1) = 0.02165
+         max over path = 0.06657  argmax α = 0.500

_sweep_logs/topk_k45_s1_mc.log

@@ -0,0 +1,43 @@
+
+mode connectivity: run=runs/topk_k45_s1  k=45  train α=[0.2, 0.4, 0.5, 0.6, 0.8]
+
+optimizing midpoint dictionary (80 steps, batch 128)...
+  step   1  meanMSE=0.05412   [0.20:0.0382 0.40:0.0623 0.50:0.0674 0.60:0.0636 0.80:0.0391]
+  step  10  meanMSE=0.04365   [0.20:0.0362 0.40:0.0474 0.50:0.0489 0.60:0.0481 0.80:0.0377]
+  step  20  meanMSE=0.04038   [0.20:0.0373 0.40:0.0472 0.50:0.0464 0.60:0.0417 0.80:0.0293]
+  step  30  meanMSE=0.04035   [0.20:0.0373 0.40:0.0464 0.50:0.0453 0.60:0.0414 0.80:0.0314]
+  step  40  meanMSE=0.04231   [0.20:0.0401 0.40:0.0482 0.50:0.0471 0.60:0.0430 0.80:0.0331]
+  step  50  meanMSE=0.04177   [0.20:0.0402 0.40:0.0487 0.50:0.0478 0.60:0.0429 0.80:0.0292]
+  step  60  meanMSE=0.04048   [0.20:0.0378 0.40:0.0461 0.50:0.0456 0.60:0.0422 0.80:0.0306]
+  step  70  meanMSE=0.03406   [0.20:0.0327 0.40:0.0394 0.50:0.0392 0.60:0.0357 0.80:0.0234]
+  step  80  meanMSE=0.04230   [0.20:0.0406 0.40:0.0468 0.50:0.0462 0.60:0.0435 0.80:0.0344]
+
+optimised in 15.7 sec
+
+evaluating optimised curve (21 αs × 5000 samples)
+  α=0.000  MSE=0.02965
+  α=0.050  MSE=0.03000
+  α=0.100  MSE=0.03170
+  α=0.150  MSE=0.03396
+  α=0.200  MSE=0.03634
+  α=0.250  MSE=0.03857
+  α=0.300  MSE=0.04046
+  α=0.350  MSE=0.04184
+  α=0.400  MSE=0.04251
+  α=0.450  MSE=0.04244
+  α=0.500  MSE=0.04160
+  α=0.550  MSE=0.04011
+  α=0.600  MSE=0.03800
+  α=0.650  MSE=0.03541
+  α=0.700  MSE=0.03256
+  α=0.750  MSE=0.02964
+  α=0.800  MSE=0.02683
+  α=0.850  MSE=0.02429
+  α=0.900  MSE=0.02207
+  α=0.950  MSE=0.02045
+  α=1.000  MSE=0.02010
+
+wrote runs/topk_k45_s1/mode_connectivity.json
+
+summary: peak=0.04251  argmax α=0.400
+         endpoints: 0.02965, 0.02010

_sweep_logs/topk_k45_s2_mc.log

@@ -0,0 +1,43 @@
+
+mode connectivity: run=runs/topk_k45_s2  k=45  train α=[0.2, 0.4, 0.5, 0.6, 0.8]
+
+optimizing midpoint dictionary (80 steps, batch 128)...
+  step   1  meanMSE=0.05019   [0.20:0.0415 0.40:0.0622 0.50:0.0635 0.60:0.0562 0.80:0.0275]
+  step  10  meanMSE=0.03718   [0.20:0.0361 0.40:0.0443 0.50:0.0429 0.60:0.0384 0.80:0.0242]
+  step  20  meanMSE=0.04184   [0.20:0.0359 0.40:0.0470 0.50:0.0472 0.60:0.0445 0.80:0.0345]
+  step  30  meanMSE=0.04091   [0.20:0.0369 0.40:0.0466 0.50:0.0463 0.60:0.0427 0.80:0.0321]
+  step  40  meanMSE=0.03408   [0.20:0.0324 0.40:0.0405 0.50:0.0396 0.60:0.0351 0.80:0.0228]
+  step  50  meanMSE=0.03986   [0.20:0.0386 0.40:0.0465 0.50:0.0452 0.60:0.0409 0.80:0.0281]
+  step  60  meanMSE=0.03756   [0.20:0.0387 0.40:0.0441 0.50:0.0423 0.60:0.0376 0.80:0.0253]
+  step  70  meanMSE=0.03755   [0.20:0.0356 0.40:0.0425 0.50:0.0417 0.60:0.0383 0.80:0.0297]
+  step  80  meanMSE=0.03776   [0.20:0.0385 0.40:0.0435 0.50:0.0423 0.60:0.0383 0.80:0.0262]
+
+optimised in 15.6 sec
+
+evaluating optimised curve (21 αs × 5000 samples)
+  α=0.000  MSE=0.02988
+  α=0.050  MSE=0.03029
+  α=0.100  MSE=0.03216
+  α=0.150  MSE=0.03468
+  α=0.200  MSE=0.03731
+  α=0.250  MSE=0.03977
+  α=0.300  MSE=0.04187
+  α=0.350  MSE=0.04334
+  α=0.400  MSE=0.04415
+  α=0.450  MSE=0.04420
+  α=0.500  MSE=0.04354
+  α=0.550  MSE=0.04225
+  α=0.600  MSE=0.04033
+  α=0.650  MSE=0.03791
+  α=0.700  MSE=0.03530
+  α=0.750  MSE=0.03261
+  α=0.800  MSE=0.03006
+  α=0.850  MSE=0.02778
+  α=0.900  MSE=0.02578
+  α=0.950  MSE=0.02422
+  α=1.000  MSE=0.02408
+
+wrote runs/topk_k45_s2/mode_connectivity.json
+
+summary: peak=0.04420  argmax α=0.450
+         endpoints: 0.02988, 0.02408

_sweep_logs/topk_k48_s2_mc.log

@@ -0,0 +1,43 @@
+
+mode connectivity: run=runs/topk_k48_s2  k=48  train α=[0.2, 0.4, 0.5, 0.6, 0.8]
+
+optimizing midpoint dictionary (80 steps, batch 128)...
+  step   1  meanMSE=0.04885   [0.20:0.0373 0.40:0.0593 0.50:0.0621 0.60:0.0562 0.80:0.0293]
+  step  10  meanMSE=0.03331   [0.20:0.0332 0.40:0.0413 0.50:0.0398 0.60:0.0346 0.80:0.0176]
+  step  20  meanMSE=0.02877   [0.20:0.0288 0.40:0.0365 0.50:0.0346 0.60:0.0296 0.80:0.0144]
+  step  30  meanMSE=0.03208   [0.20:0.0326 0.40:0.0393 0.50:0.0375 0.60:0.0328 0.80:0.0182]
+  step  40  meanMSE=0.03851   [0.20:0.0376 0.40:0.0461 0.50:0.0448 0.60:0.0395 0.80:0.0246]
+  step  50  meanMSE=0.03263   [0.20:0.0313 0.40:0.0386 0.50:0.0377 0.60:0.0339 0.80:0.0216]
+  step  60  meanMSE=0.03791   [0.20:0.0366 0.40:0.0436 0.50:0.0430 0.60:0.0394 0.80:0.0270]
+  step  70  meanMSE=0.03371   [0.20:0.0348 0.40:0.0404 0.50:0.0388 0.60:0.0343 0.80:0.0202]
+  step  80  meanMSE=0.03407   [0.20:0.0333 0.40:0.0400 0.50:0.0393 0.60:0.0356 0.80:0.0222]
+
+optimised in 16.9 sec
+
+evaluating optimised curve (21 αs × 5000 samples)
+  α=0.000  MSE=0.02708
+  α=0.050  MSE=0.02740
+  α=0.100  MSE=0.02909
+  α=0.150  MSE=0.03136
+  α=0.200  MSE=0.03374
+  α=0.250  MSE=0.03599
+  α=0.300  MSE=0.03788
+  α=0.350  MSE=0.03925
+  α=0.400  MSE=0.03999
+  α=0.450  MSE=0.03998
+  α=0.500  MSE=0.03920
+  α=0.550  MSE=0.03772
+  α=0.600  MSE=0.03554
+  α=0.650  MSE=0.03279
+  α=0.700  MSE=0.02966
+  α=0.750  MSE=0.02631
+  α=0.800  MSE=0.02289
+  α=0.850  MSE=0.01957
+  α=0.900  MSE=0.01650
+  α=0.950  MSE=0.01408
+  α=1.000  MSE=0.01338
+
+wrote runs/topk_k48_s2/mode_connectivity.json
+
+summary: peak=0.03999  argmax α=0.400
+         endpoints: 0.02708, 0.01338

_sweep_logs/topk_k50_s1_barrier.log

@@ -0,0 +1,30 @@
+
+barrier eval: run=runs/topk_k50_s1  k=50  n=5000  alphas=21
+
+Hungarian alignment: mean cos = 0.7114 (0.7114 per scipy)
+  α=0.000  MSE=0.02412
+  α=0.050  MSE=0.02465
+  α=0.100  MSE=0.02638
+  α=0.150  MSE=0.02917
+  α=0.200  MSE=0.03299
+  α=0.250  MSE=0.03777
+  α=0.300  MSE=0.04328
+  α=0.350  MSE=0.04895
+  α=0.400  MSE=0.05391
+  α=0.450  MSE=0.05709
+  α=0.500  MSE=0.05774
+  α=0.550  MSE=0.05586
+  α=0.600  MSE=0.05166
+  α=0.650  MSE=0.04569
+  α=0.700  MSE=0.03836
+  α=0.750  MSE=0.03068
+  α=0.800  MSE=0.02352
+  α=0.850  MSE=0.01745
+  α=0.900  MSE=0.01289
+  α=0.950  MSE=0.01013
+  α=1.000  MSE=0.00932
+
+wrote runs/topk_k50_s1/barrier.json
+
+summary: MSE(α=0) = 0.02412  ->  MSE(α=1) = 0.00932
+         max over path = 0.05774  argmax α = 0.500

_sweep_logs/topk_k50_s1_mc.log

@@ -0,0 +1,43 @@
+
+mode connectivity: run=runs/topk_k50_s1  k=50  train α=[0.2, 0.4, 0.5, 0.6, 0.8]
+
+optimizing midpoint dictionary (80 steps, batch 128)...
+  step   1  meanMSE=0.04677   [0.20:0.0357 0.40:0.0571 0.50:0.0609 0.60:0.0543 0.80:0.0259]
+  step  10  meanMSE=0.02857   [0.20:0.0274 0.40:0.0351 0.50:0.0343 0.60:0.0301 0.80:0.0160]
+  step  20  meanMSE=0.02857   [0.20:0.0291 0.40:0.0366 0.50:0.0344 0.60:0.0292 0.80:0.0135]
+  step  30  meanMSE=0.03326   [0.20:0.0298 0.40:0.0391 0.50:0.0389 0.60:0.0357 0.80:0.0229]
+  step  40  meanMSE=0.03114   [0.20:0.0318 0.40:0.0389 0.50:0.0369 0.60:0.0318 0.80:0.0164]
+  step  50  meanMSE=0.03590   [0.20:0.0338 0.40:0.0428 0.50:0.0420 0.60:0.0376 0.80:0.0233]
+  step  60  meanMSE=0.03094   [0.20:0.0300 0.40:0.0372 0.50:0.0365 0.60:0.0326 0.80:0.0184]
+  step  70  meanMSE=0.03008   [0.20:0.0301 0.40:0.0367 0.50:0.0358 0.60:0.0316 0.80:0.0162]
+  step  80  meanMSE=0.02948   [0.20:0.0270 0.40:0.0348 0.50:0.0349 0.60:0.0317 0.80:0.0190]
+
+optimised in 18.1 sec
+
+evaluating optimised curve (21 αs × 5000 samples)
+  α=0.000  MSE=0.02413
+  α=0.050  MSE=0.02446
+  α=0.100  MSE=0.02617
+  α=0.150  MSE=0.02854
+  α=0.200  MSE=0.03113
+  α=0.250  MSE=0.03361
+  α=0.300  MSE=0.03572
+  α=0.350  MSE=0.03732
+  α=0.400  MSE=0.03828
+  α=0.450  MSE=0.03846
+  α=0.500  MSE=0.03783
+  α=0.550  MSE=0.03636
+  α=0.600  MSE=0.03418
+  α=0.650  MSE=0.03141
+  α=0.700  MSE=0.02818
+  α=0.750  MSE=0.02462
+  α=0.800  MSE=0.02094
+  α=0.850  MSE=0.01723
+  α=0.900  MSE=0.01385
+  α=0.950  MSE=0.01122
+  α=1.000  MSE=0.01036
+
+wrote runs/topk_k50_s1/mode_connectivity.json
+
+summary: peak=0.03846  argmax α=0.450
+         endpoints: 0.02413, 0.01036

_sweep_logs/topk_k50_s2_barrier.log

@@ -0,0 +1,30 @@
+
+barrier eval: run=runs/topk_k50_s2  k=50  n=5000  alphas=21
+
+Hungarian alignment: mean cos = 0.7091 (0.7091 per scipy)
+  α=0.000  MSE=0.02492
+  α=0.050  MSE=0.02552
+  α=0.100  MSE=0.02761
+  α=0.150  MSE=0.03107
+  α=0.200  MSE=0.03573
+  α=0.250  MSE=0.04131
+  α=0.300  MSE=0.04735
+  α=0.350  MSE=0.05317
+  α=0.400  MSE=0.05783
+  α=0.450  MSE=0.06035
+  α=0.500  MSE=0.06025
+  α=0.550  MSE=0.05780
+  α=0.600  MSE=0.05331
+  α=0.650  MSE=0.04725
+  α=0.700  MSE=0.03984
+  α=0.750  MSE=0.03205
+  α=0.800  MSE=0.02483
+  α=0.850  MSE=0.01874
+  α=0.900  MSE=0.01422
+  α=0.950  MSE=0.01150
+  α=1.000  MSE=0.01075
+
+wrote runs/topk_k50_s2/barrier.json
+
+summary: MSE(α=0) = 0.02492  ->  MSE(α=1) = 0.01075
+         max over path = 0.06035  argmax α = 0.450

altmin_perturbed_omp_k48_s43_c0.30/results.json

@@ -0,0 +1,23 @@
+{
+  "init": "perturbed",
+  "coding": "omp",
+  "k": 48,
+  "n_samples": 10000000,
+  "lr": 0.0003,
+  "seed": 43,
+  "eval": {
+    "true_l0": 34.493215,
+    "sae_l0": 38.94516,
+    "dead_latents": 0,
+    "shrinkage": 0.9630266931152344,
+    "explained_variance": 0.9091709596347162,
+    "mcc": 0.9172072410583496,
+    "uniqueness": 0.9921875,
+    "classification/precision": 0.6697086691856384,
+    "classification/recall": 0.6960408687591553,
+    "classification/f1_score": 0.6599588394165039,
+    "classification/accuracy": 0.9984580278396606
+  },
+  "dec_fidelity_final": 0.9184098839759827,
+  "seconds": 3053.9301381111145
+}

altmin_perturbed_omp_k48_s43_c0.40/results.json

@@ -0,0 +1,23 @@
+{
+  "init": "perturbed",
+  "coding": "omp",
+  "k": 48,
+  "n_samples": 10000000,
+  "lr": 0.0003,
+  "seed": 43,
+  "eval": {
+    "true_l0": 34.493215,
+    "sae_l0": 39.43884,
+    "dead_latents": 0,
+    "shrinkage": 0.9721412866210938,
+    "explained_variance": 0.9291685954011948,
+    "mcc": 0.9540121555328369,
+    "uniqueness": 0.99951171875,
+    "classification/precision": 0.7348754405975342,
+    "classification/recall": 0.750920832157135,
+    "classification/f1_score": 0.7241045236587524,
+    "classification/accuracy": 0.9988424777984619
+  },
+  "dec_fidelity_final": 0.9541122913360596,
+  "seconds": 3065.881599664688
+}

basin_radius/results_seed42.json

@@ -0,0 +1,104 @@
+{
+  "k": 48,
+  "n_steps": 500,
+  "n_samples": 2048000,
+  "seed": 42,
+  "results": [
+    {
+      "target_cos": 0.99,
+      "fid_init": 0.9900000095367432,
+      "fid_final": 0.9861536026000977,
+      "delta": -0.003846406936645508,
+      "converging": false,
+      "seconds": 448.9438157081604
+    },
+    {
+      "target_cos": 0.97,
+      "fid_init": 0.9700000286102295,
+      "fid_final": 0.9848216772079468,
+      "delta": 0.014821648597717285,
+      "converging": true,
+      "seconds": 452.4573805332184
+    },
+    {
+      "target_cos": 0.95,
+      "fid_init": 0.9500000476837158,
+      "fid_final": 0.9826305508613586,
+      "delta": 0.03263050317764282,
+      "converging": true,
+      "seconds": 453.744690656662
+    },
+    {
+      "target_cos": 0.9,
+      "fid_init": 0.9000000953674316,
+      "fid_final": 0.9756519198417664,
+      "delta": 0.07565182447433472,
+      "converging": true,
+      "seconds": 455.57620120048523
+    },
+    {
+      "target_cos": 0.85,
+      "fid_init": 0.8500000238418579,
+      "fid_final": 0.9676114320755005,
+      "delta": 0.11761140823364258,
+      "converging": true,
+      "seconds": 456.31761837005615
+    },
+    {
+      "target_cos": 0.8,
+      "fid_init": 0.8000000715255737,
+      "fid_final": 0.9589360356330872,
+      "delta": 0.15893596410751343,
+      "converging": true,
+      "seconds": 455.580677986145
+    },
+    {
+      "target_cos": 0.7,
+      "fid_init": 0.7118338346481323,
+      "fid_final": 0.8864246010780334,
+      "delta": 0.17459076642990112,
+      "converging": true,
+      "seconds": 455.87197256088257
+    },
+    {
+      "target_cos": 0.6,
+      "fid_init": 0.7552973031997681,
+      "fid_final": 0.866296648979187,
+      "delta": 0.11099934577941895,
+      "converging": true,
+      "seconds": 455.4593770503998
+    },
+    {
+      "target_cos": 0.5,
+      "fid_init": 0.7880702614784241,
+      "fid_final": 0.8552908897399902,
+      "delta": 0.06722062826156616,
+      "converging": true,
+      "seconds": 455.7541358470917
+    },
+    {
+      "target_cos": 0.3,
+      "fid_init": 0.8184905052185059,
+      "fid_final": 0.8534693717956543,
+      "delta": 0.03497886657714844,
+      "converging": true,
+      "seconds": 456.12142276763916
+    },
+    {
+      "target_cos": 0.2,
+      "fid_init": 0.8192972540855408,
+      "fid_final": 0.8383699655532837,
+      "delta": 0.01907271146774292,
+      "converging": true,
+      "seconds": 455.7897198200226
+    },
+    {
+      "target_cos": 0.1,
+      "fid_init": 0.8211536407470703,
+      "fid_final": 0.8282573223114014,
+      "delta": 0.007103681564331055,
+      "converging": true,
+      "seconds": 455.4574017524719
+    }
+  ]
+}

basin_radius/results_seed43.json

@@ -0,0 +1,104 @@
+{
+  "k": 48,
+  "n_steps": 500,
+  "n_samples": 2048000,
+  "seed": 43,
+  "results": [
+    {
+      "target_cos": 0.99,
+      "fid_init": 0.9900000095367432,
+      "fid_final": 0.9860057830810547,
+      "delta": -0.0039942264556884766,
+      "converging": false,
+      "seconds": 448.4635183811188
+    },
+    {
+      "target_cos": 0.97,
+      "fid_init": 0.9700000286102295,
+      "fid_final": 0.9846746921539307,
+      "delta": 0.014674663543701172,
+      "converging": true,
+      "seconds": 447.87573170661926
+    },
+    {
+      "target_cos": 0.95,
+      "fid_init": 0.9500000476837158,
+      "fid_final": 0.9829118251800537,
+      "delta": 0.03291177749633789,
+      "converging": true,
+      "seconds": 448.0157024860382
+    },
+    {
+      "target_cos": 0.9,
+      "fid_init": 0.9000000953674316,
+      "fid_final": 0.9769527316093445,
+      "delta": 0.07695263624191284,
+      "converging": true,
+      "seconds": 448.6988363265991
+    },
+    {
+      "target_cos": 0.85,
+      "fid_init": 0.8500000238418579,
+      "fid_final": 0.969549298286438,
+      "delta": 0.11954927444458008,
+      "converging": true,
+      "seconds": 448.3461797237396
+    },
+    {
+      "target_cos": 0.8,
+      "fid_init": 0.8000000715255737,
+      "fid_final": 0.9612268209457397,
+      "delta": 0.16122674942016602,
+      "converging": true,
+      "seconds": 448.9183044433594
+    },
+    {
+      "target_cos": 0.7,
+      "fid_init": 0.7000000476837158,
+      "fid_final": 0.9416828155517578,
+      "delta": 0.241682767868042,
+      "converging": true,
+      "seconds": 449.0006408691406
+    },
+    {
+      "target_cos": 0.6,
+      "fid_init": 0.6000000238418579,
+      "fid_final": 0.9099850654602051,
+      "delta": 0.30998504161834717,
+      "converging": true,
+      "seconds": 449.1651108264923
+    },
+    {
+      "target_cos": 0.5,
+      "fid_init": 0.5,
+      "fid_final": 0.8409985303878784,
+      "delta": 0.3409985303878784,
+      "converging": true,
+      "seconds": 450.0478971004486
+    },
+    {
+      "target_cos": 0.3,
+      "fid_init": 0.30000001192092896,
+      "fid_final": 0.41315793991088867,
+      "delta": 0.11315792798995972,
+      "converging": true,
+      "seconds": 448.98216104507446
+    },
+    {
+      "target_cos": 0.2,
+      "fid_init": 0.20000237226486206,
+      "fid_final": 0.2386893630027771,
+      "delta": 0.03868699073791504,
+      "converging": true,
+      "seconds": 449.53068566322327
+    },
+    {
+      "target_cos": 0.1,
+      "fid_init": 0.14860697090625763,
+      "fid_final": 0.21956323087215424,
+      "delta": 0.0709562599658966,
+      "converging": true,
+      "seconds": 450.1033709049225
+    }
+  ]
+}

basin_radius/results_seed44.json

@@ -0,0 +1,104 @@
+{
+  "k": 48,
+  "n_steps": 500,
+  "n_samples": 2048000,
+  "seed": 44,
+  "results": [
+    {
+      "target_cos": 0.99,
+      "fid_init": 0.9900000095367432,
+      "fid_final": 0.9861334562301636,
+      "delta": -0.00386655330657959,
+      "converging": false,
+      "seconds": 458.3089339733124
+    },
+    {
+      "target_cos": 0.97,
+      "fid_init": 0.9700000286102295,
+      "fid_final": 0.9848356246948242,
+      "delta": 0.014835596084594727,
+      "converging": true,
+      "seconds": 461.82796478271484
+    },
+    {
+      "target_cos": 0.95,
+      "fid_init": 0.9500000476837158,
+      "fid_final": 0.983006477355957,
+      "delta": 0.03300642967224121,
+      "converging": true,
+      "seconds": 461.8896379470825
+    },
+    {
+      "target_cos": 0.9,
+      "fid_init": 0.9000000953674316,
+      "fid_final": 0.9769294261932373,
+      "delta": 0.07692933082580566,
+      "converging": true,
+      "seconds": 460.7876327037811
+    },
+    {
+      "target_cos": 0.85,
+      "fid_init": 0.8500000238418579,
+      "fid_final": 0.9694505333900452,
+      "delta": 0.11945050954818726,
+      "converging": true,
+      "seconds": 461.25571632385254
+    },
+    {
+      "target_cos": 0.8,
+      "fid_init": 0.8000000715255737,
+      "fid_final": 0.9610999822616577,
+      "delta": 0.16109991073608398,
+      "converging": true,
+      "seconds": 465.52943754196167
+    },
+    {
+      "target_cos": 0.7,
+      "fid_init": 0.7000000476837158,
+      "fid_final": 0.9415371417999268,
+      "delta": 0.24153709411621094,
+      "converging": true,
+      "seconds": 461.8532736301422
+    },
+    {
+      "target_cos": 0.6,
+      "fid_init": 0.6000000238418579,
+      "fid_final": 0.9099147319793701,
+      "delta": 0.3099147081375122,
+      "converging": true,
+      "seconds": 463.27074575424194
+    },
+    {
+      "target_cos": 0.5,
+      "fid_init": 0.5,
+      "fid_final": 0.8409147262573242,
+      "delta": 0.3409147262573242,
+      "converging": true,
+      "seconds": 464.4487407207489
+    },
+    {
+      "target_cos": 0.3,
+      "fid_init": 0.30000001192092896,
+      "fid_final": 0.4143797755241394,
+      "delta": 0.11437976360321045,
+      "converging": true,
+      "seconds": 463.4620110988617
+    },
+    {
+      "target_cos": 0.2,
+      "fid_init": 0.20000238716602325,
+      "fid_final": 0.24059367179870605,
+      "delta": 0.0405912846326828,
+      "converging": true,
+      "seconds": 457.25708651542664
+    },
+    {
+      "target_cos": 0.1,
+      "fid_init": 0.14846540987491608,
+      "fid_final": 0.22045353055000305,
+      "delta": 0.07198812067508698,
+      "converging": true,
+      "seconds": 459.9436514377594
+    }
+  ]
+}

frozen_dec_k48_s1/results.json

@@ -0,0 +1,26 @@
+{
+  "k": 48,
+  "n_samples": 15000000,
+  "lr": 0.0003,
+  "seed": 1,
+  "support_overlap_init": {
+    "jaccard": 0.4592665173012477,
+    "recall": 0.6161499023437506
+  },
+  "support_overlap_final": {
+    "jaccard": 0.5475213736319267,
+    "recall": 0.6813151041666667
+  },
+  "dec_fidelity_init": 1.0,
+  "dec_fidelity_final": 1.0,
+  "eval": {
+    "true_l0": 34.46586,
+    "sae_l0": 47.98879,
+    "dead_latents": 0,
+    "shrinkage": 0.9727815844726563,
+    "explained_variance": 0.9287085583622611,
+    "mcc": 1.0,
+    "uniqueness": 1.0
+  },
+  "seconds": 160.0245246887207
+}

oracle_init_smoke/cfg.json

@@ -0,0 +1 @@
+{"normalize_activations": "none", "dtype": "float32", "device": "cuda", "d_in": 768, "rescale_acts_by_decoder_norm": true, "apply_b_dec_to_input": true, "d_sae": 16384, "k": 48, "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "reshape_activations": "none", "architecture": "topk"}

oracle_init_smoke/ckpts/114688/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/114688/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/114688/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/225280/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/225280/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/225280/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/335872/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/335872/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/335872/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/446464/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/446464/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/446464/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/557056/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/557056/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/557056/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/667648/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/667648/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/667648/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/778240/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/778240/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/778240/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/892928/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": 0.9888814189820523}

oracle_init_smoke/ckpts/892928/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "expected_average_only_in", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/892928/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "expected_average_only_in",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/ckpts/final_1003520/activation_scaler.json

@@ -0,0 +1 @@
+{"scaling_factor": null}

oracle_init_smoke/ckpts/final_1003520/cfg.json

@@ -0,0 +1 @@
+{"d_in": 768, "d_sae": 16384, "dtype": "float32", "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "none", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.43.0", "sae_lens_training_version": "6.43.0"}, "decoder_init_norm": 0.1, "k": 48, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "architecture": "topk"}

oracle_init_smoke/ckpts/final_1003520/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "none",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}

oracle_init_smoke/eval_stats.json

@@ -0,0 +1,15 @@
+{
+  "true_l0": 34.46811,
+  "sae_l0": 47.999505,
+  "dead_latents": 0,
+  "shrinkage": 0.9734420861816406,
+  "explained_variance": 0.9196613902159156,
+  "mcc": 0.990073561668396,
+  "uniqueness": 1.0,
+  "classification": {
+    "precision": 0.5700647234916687,
+    "recall": 0.91277015209198,
+    "f1_score": 0.6972928047180176,
+    "accuracy": 0.998992383480072
+  }
+}

oracle_init_smoke/meta.json

@@ -0,0 +1,8 @@
+{
+  "k": 48,
+  "seed": 1,
+  "n_samples": 1000000,
+  "cos_init": 1.0,
+  "cos_final": 0.990073561668396,
+  "seconds": 1409.8078198432922
+}

oracle_init_smoke/runner_config.json

@@ -0,0 +1,59 @@
+{
+  "synthetic_model": "decoderesearch/synth-sae-bench-16k-v1",
+  "sae": {
+    "d_in": 768,
+    "d_sae": 16384,
+    "dtype": "float32",
+    "device": "cuda",
+    "apply_b_dec_to_input": true,
+    "normalize_activations": "none",
+    "reshape_activations": "none",
+    "metadata": {
+      "sae_lens_version": "6.43.0",
+      "sae_lens_training_version": "6.43.0"
+    },
+    "decoder_init_norm": 0.1,
+    "k": 48,
+    "use_sparse_activations": false,
+    "aux_loss_coefficient": 1.0,
+    "rescale_acts_by_decoder_norm": true,
+    "architecture": "topk"
+  },
+  "training_samples": 1000000,
+  "batch_size": 4096,
+  "lr": 0.0003,
+  "lr_warm_up_steps": 0,
+  "lr_decay_steps": 0,
+  "lr_scheduler_name": "constant",
+  "lr_end": 2.9999999999999997e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "n_restart_cycles": 1,
+  "device": "cuda",
+  "autocast_sae": false,
+  "autocast_data": false,
+  "n_checkpoints": 8,
+  "checkpoint_path": "runs/oracle_init_smoke/ckpts",
+  "save_final_checkpoint": true,
+  "output_path": "runs/oracle_init_smoke",
+  "save_synthetic_model": false,
+  "eval_frequency": 200000,
+  "eval_samples": 200000,
+  "run_final_eval": true,
+  "dead_feature_window": 1000,
+  "feature_sampling_window": 2000,
+  "n_batches_for_norm_estimate": 1000,
+  "sae_lens_version": "6.43.0",
+  "logger": {
+    "log_to_wandb": true,
+    "log_activations_store_to_wandb": false,
+    "log_optimizer_state_to_wandb": false,
+    "log_weights_to_wandb": true,
+    "wandb_project": "sae_lens_training",
+    "wandb_id": null,
+    "run_name": "synthetic-topk-16384-LR-0.0003",
+    "wandb_entity": null,
+    "wandb_log_frequency": 10,
+    "eval_every_n_wandb_logs": 100
+  }
+}