Dataset Viewer
Auto-converted to Parquet Duplicate
family
stringclasses
10 values
model_ref
stringlengths
38
68
step
int64
0
126k
tokens
float64
brain_rsa_mean
float64
-0.04
0.02
brain_rsa_std
float64
0.01
0.06
brain_rsa_pearson_mean
float64
-0.03
0.01
brain_n_cells
int64
12
12
brain_rsa_Sem
float64
-0.1
0.05
brain_rsa_Phon
float64
-0.06
0.11
brain_rsa_Gram
float64
-0.04
0.02
brain_rsa_Plaus
float64
-0.08
0.05
interp_norm
float64
7.87
782
interp_gini
float64
0.06
0.36
interp_hoyer
float64
0.01
0.79
interp_per
float64
0.01
0.39
interp_condition_number
float64
17.1
774
interp_cka_to_prev
float64
0.15
1
loc_selectivity
float64
0.43
0.59
loc_overlap
float64
0
0.03
loc_gini
float64
0.36
0.42
loc_entropy
float64
0.96
0.98
loc_layer_com
float64
0.34
0.68
loc_n_active_layers
float64
6.25
14
behav_mp_accuracy
float64
0.43
0.77
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-01
598
null
-0.024803
0.030126
-0.016878
12
0.010718
-0.015072
-0.02992
-0.064935
52.742121
0.31721
0.152533
0.049836
376.864096
null
0.523755
0.003653
0.415383
0.968718
0.512152
12
0.677083
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-02
897
null
-0.039048
0.03626
-0.023949
12
0.007038
-0.055822
-0.031703
-0.075705
56.075028
0.317426
0.161777
0.06074
272.553093
0.949594
0.535215
0.001821
0.417779
0.968367
0.45972
12
0.6625
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-03
1,196
null
-0.027308
0.025661
-0.020222
12
0.00693
-0.029605
-0.036011
-0.050547
56.396128
0.297967
0.166181
0.076489
213.573084
0.941563
0.521009
0.008329
0.412421
0.969084
0.476405
12
0.689583
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-04
1,495
null
-0.036588
0.023906
-0.024623
12
-0.008144
-0.042021
-0.03329
-0.062898
52.494003
0.289181
0.172938
0.074367
187.440928
0.968135
0.542142
0.001821
0.415326
0.9687
0.482922
12
0.7125
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-05
1,794
null
-0.035533
0.021915
-0.025109
12
-0.006481
-0.052915
-0.036845
-0.04589
53.486169
0.288556
0.179965
0.080345
174.908934
0.973153
0.529266
0.003668
0.414982
0.968793
0.52024
12
0.74375
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-06
2,093
null
-0.032252
0.01895
-0.024268
12
-0.006167
-0.040721
-0.033843
-0.048278
51.218507
0.283115
0.190178
0.074221
172.458474
0.982936
0.525783
0.006405
0.413312
0.969024
0.506209
12
0.722917
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-07
2,392
null
-0.031539
0.023419
-0.023895
12
-0.000902
-0.042995
-0.031602
-0.050655
49.521919
0.280993
0.19228
0.076248
163.63048
0.992566
0.521275
0.004574
0.412564
0.969134
0.53408
12
0.7
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-08
2,691
null
-0.031626
0.023258
-0.023628
12
0.000303
-0.042129
-0.031885
-0.052791
48.219034
0.282759
0.19323
0.078539
162.043283
0.997007
0.535448
0.003653
0.412753
0.96911
0.529375
12
0.720833
babylm-gpt2
BrainAlign/gpt2-babylm-9@checkpoint-09
2,990
null
-0.032054
0.022575
-0.023756
12
-0.00192
-0.040914
-0.031574
-0.053806
48.148181
0.282287
0.195198
0.078349
161.377812
0.999652
0.537871
0.003653
0.414016
0.968907
0.535451
12
0.73125
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-01
191
null
-0.021981
0.035991
-0.011153
12
0.022254
-0.028775
-0.025432
-0.055973
83.710765
0.348298
0.165259
0.049512
604.084537
null
0.507756
0.009482
0.407918
0.969781
0.494014
12
0.56875
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-02
382
null
-0.029955
0.031823
-0.022239
12
0.014082
-0.052239
-0.023362
-0.0583
72.206114
0.359109
0.176159
0.04827
567.742516
0.817417
0.53777
0.002742
0.412393
0.969143
0.53233
12
0.7
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-03
573
null
-0.028228
0.038556
-0.024405
12
0.021614
-0.043977
-0.021519
-0.069029
61.712115
0.338975
0.163855
0.05054
435.205883
0.938142
0.521496
0.002762
0.413007
0.969079
0.519653
12
0.7125
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-04
764
null
-0.02729
0.036081
-0.02043
12
0.018277
-0.03705
-0.022799
-0.06759
64.601577
0.343441
0.172404
0.058871
360.532611
0.959111
0.49681
0.007398
0.412668
0.969116
0.484946
12
0.641667
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-05
1,146
null
-0.021642
0.033671
-0.015351
12
0.020728
-0.01824
-0.026476
-0.062582
53.726943
0.318056
0.15175
0.077934
255.845989
0.954175
0.517571
0.013155
0.413804
0.968961
0.55087
12
0.75625
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-06
1,337
null
-0.018489
0.035472
-0.013356
12
0.028242
-0.018428
-0.025416
-0.058356
50.034724
0.311943
0.148518
0.072302
254.435631
0.983721
0.527811
0.010393
0.415759
0.968659
0.516136
12
0.689583
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-07
1,528
null
-0.014787
0.028666
-0.011
12
0.022732
-0.012547
-0.026165
-0.043167
49.803259
0.306378
0.147007
0.07575
243.742111
0.991501
0.525316
0.008267
0.416531
0.968541
0.510021
12
0.716667
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-08
1,719
null
-0.012376
0.028795
-0.008227
12
0.024751
-0.008509
-0.024203
-0.041543
47.891462
0.303679
0.145224
0.077354
233.725258
0.997233
0.517285
0.007346
0.41606
0.968632
0.501055
12
0.739583
babylm-gpt2-3
BrainAlign/gpt2-babylm-3@checkpoint-09
1,908
null
-0.013295
0.029954
-0.009212
12
0.024573
-0.009117
-0.023512
-0.045123
47.757338
0.304427
0.145683
0.076861
233.124327
0.999603
0.520274
0.008329
0.416502
0.968579
0.511387
12
0.714583
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-01
191
null
-0.023959
0.034402
-0.012289
12
0.018571
-0.026546
-0.030278
-0.057582
83.533602
0.352538
0.167748
0.047343
616.141456
null
0.499769
0.010489
0.408954
0.969602
0.512294
12
0.58125
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-02
382
null
-0.028338
0.031817
-0.021266
12
0.014883
-0.056109
-0.021735
-0.050393
72.555552
0.351489
0.168743
0.048909
554.573751
0.834237
0.538939
0.00638
0.41053
0.969458
0.512667
12
0.65625
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-03
573
null
-0.025382
0.037411
-0.022331
12
0.02485
-0.036568
-0.026065
-0.063747
61.930485
0.338525
0.164011
0.050929
415.195292
0.936611
0.547796
0.002742
0.412817
0.969109
0.502404
12
0.6875
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-04
764
null
-0.025882
0.035222
-0.019802
12
0.020271
-0.040295
-0.020531
-0.062974
64.80808
0.343997
0.171882
0.058776
354.43068
0.958418
0.516966
0.005505
0.411858
0.969275
0.477051
12
0.641667
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-05
1,146
null
-0.023841
0.033881
-0.017145
12
0.018403
-0.019102
-0.028562
-0.066101
54.089538
0.316283
0.151575
0.079086
267.505065
0.954206
0.535047
0.005505
0.415443
0.968723
0.535684
12
0.74375
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-06
1,337
null
-0.019022
0.033021
-0.014119
12
0.024473
-0.022584
-0.022559
-0.055419
49.546049
0.306677
0.145933
0.076004
258.041043
0.984051
0.525096
0.009168
0.414815
0.968762
0.54472
12
0.702083
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-07
1,528
null
-0.016324
0.02757
-0.012021
12
0.018847
-0.013194
-0.02628
-0.044667
49.905949
0.301789
0.143019
0.077339
243.614897
0.991509
0.541681
0.005525
0.416121
0.968624
0.509992
12
0.714583
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-08
1,719
null
-0.01532
0.028372
-0.010961
12
0.02103
-0.011731
-0.025525
-0.045055
48.271498
0.299628
0.142336
0.077194
240.066065
0.997346
0.531831
0.004564
0.416588
0.96853
0.494547
12
0.741667
babylm-gpt2-5
BrainAlign/gpt2-babylm-5@checkpoint-09
1,908
null
-0.016244
0.028951
-0.011664
12
0.020931
-0.013782
-0.02563
-0.046494
48.117843
0.300808
0.143075
0.077533
238.442461
0.999641
0.539417
0.003653
0.417575
0.968401
0.503275
12
0.716667
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-01
191
null
-0.024374
0.035624
-0.012405
12
0.021201
-0.034667
-0.027736
-0.056296
86.046775
0.350655
0.166937
0.046899
634.253303
null
0.492891
0.008571
0.407529
0.969915
0.50044
12
0.58125
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-02
382
null
-0.029043
0.032293
-0.02249
12
0.015243
-0.052586
-0.021222
-0.057608
73.92994
0.354578
0.171399
0.048754
562.972299
0.822586
0.554597
0.003653
0.411086
0.969303
0.533863
12
0.7
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-03
573
null
-0.026382
0.034363
-0.025099
12
0.019106
-0.0336
-0.027705
-0.06333
61.961444
0.341751
0.165982
0.055294
402.214186
0.935793
0.534018
0.005535
0.415126
0.968751
0.52242
12
0.689583
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-04
764
null
-0.028042
0.036459
-0.02062
12
0.016129
-0.036332
-0.022311
-0.069653
65.597931
0.345055
0.171958
0.056403
370.905045
0.956719
0.514293
0.006436
0.415929
0.968608
0.513651
12
0.689583
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-05
1,146
null
-0.025204
0.028874
-0.018815
12
0.009749
-0.021452
-0.029549
-0.059565
56.435083
0.323637
0.157489
0.074073
267.993869
0.956426
0.512567
0.004614
0.416569
0.968529
0.520626
12
0.766667
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-06
1,337
null
-0.022117
0.0306
-0.016235
12
0.016757
-0.024709
-0.025142
-0.055375
50.504937
0.309422
0.147618
0.071154
264.118807
0.981446
0.505202
0.014858
0.414893
0.968783
0.539158
12
0.727083
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-07
1,528
null
-0.019761
0.023722
-0.014869
12
0.010178
-0.019215
-0.026818
-0.04319
51.319222
0.31099
0.151211
0.076494
246.351735
0.988197
0.527835
0.005474
0.416734
0.968481
0.506433
12
0.73125
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-08
1,719
null
-0.018103
0.024744
-0.012581
12
0.013256
-0.016723
-0.025332
-0.043612
48.871656
0.305345
0.147609
0.077112
238.53276
0.99736
0.516178
0.005495
0.415675
0.968647
0.502766
12
0.741667
babylm-gpt2-7
BrainAlign/gpt2-babylm-7@checkpoint-09
1,908
null
-0.018625
0.025156
-0.013048
12
0.013106
-0.018064
-0.025243
-0.044297
48.946432
0.307833
0.149525
0.076909
238.000177
0.999584
0.517597
0.005484
0.416543
0.968553
0.500877
12
0.73125
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-0
0
null
-0.0156
0.019411
-0.010195
12
-0.008507
0.007931
-0.036482
-0.025345
22.902792
0.061261
0.006236
0.300869
21.861143
null
0.566278
0.001572
0.417842
0.968925
0.54588
14
0.558333
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-1
1
null
-0.0156
0.019411
-0.010195
12
-0.008507
0.007931
-0.036482
-0.025345
22.902792
0.061261
0.006236
0.300869
21.861143
1
0.566278
0.001572
0.417842
0.968925
0.54588
14
0.558333
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-2
2
null
-0.0156
0.019411
-0.010195
12
-0.008507
0.007931
-0.036482
-0.025345
22.902792
0.061261
0.006236
0.300869
21.861143
1
0.566278
0.001572
0.417842
0.968925
0.54588
14
0.558333
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-4
4
null
-0.015601
0.019419
-0.010198
12
-0.008534
0.007961
-0.036494
-0.025335
22.90277
0.061261
0.006236
0.300869
21.861358
1
0.566277
0.001572
0.417845
0.968924
0.545878
14
0.558333
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-6
6
null
-0.015591
0.0194
-0.010201
12
-0.008587
0.007983
-0.036423
-0.025338
22.902757
0.061261
0.006236
0.300866
21.862151
1
0.566283
0.001572
0.417853
0.968923
0.545876
14
0.558333
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-8
8
null
-0.015609
0.019392
-0.010202
12
-0.008589
0.007931
-0.036469
-0.025311
22.902723
0.061259
0.006235
0.300865
21.86307
1
0.566278
0.001572
0.417862
0.968921
0.545871
14
0.547917
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-16
16
null
-0.015544
0.01941
-0.010198
12
-0.008601
0.008059
-0.036465
-0.025167
22.902691
0.061249
0.006233
0.300855
21.870342
0.999996
0.566957
0.001572
0.417895
0.968913
0.545997
14
0.5375
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-32
32
null
-0.01547
0.019687
-0.010204
12
-0.009376
0.008955
-0.036746
-0.024713
22.907808
0.061228
0.006231
0.300751
21.896785
0.999919
0.562846
0.001572
0.417882
0.968907
0.544457
14
0.5375
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-64
64
null
-0.010948
0.022347
-0.008891
12
-0.016904
0.016672
-0.039448
-0.004113
23.395499
0.064455
0.007044
0.273118
24.73127
0.949357
0.572038
0.001565
0.417753
0.968927
0.549516
14
0.535417
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-128
128
null
-0.005405
0.018775
-0.00129
12
-0.008821
0.006055
-0.031309
0.012453
30.282229
0.115911
0.02778
0.174359
58.101352
0.824128
0.558138
0.001565
0.416487
0.969097
0.499484
14
0.579167
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-256
256
null
-0.003812
0.039473
0.011212
12
0.025275
0.022143
-0.005086
-0.057582
31.409324
0.101536
0.02011
0.232598
37.666329
0.741334
0.561615
0
0.414833
0.969351
0.544163
14
0.585417
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-512
512
null
-0.007462
0.020472
0.000093
12
0.001729
-0.00307
-0.030663
0.002157
28.967892
0.090613
0.015472
0.240539
30.791498
0.766808
0.574661
0
0.414751
0.969366
0.555575
14
0.577083
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-1024
1,024
null
-0.006002
0.021632
-0.007479
12
0.009751
0.013398
-0.032099
-0.015061
33.248203
0.086436
0.014056
0.213362
39.859558
0.826383
0.541584
0.00392
0.41395
0.969417
0.502089
14
0.5875
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-2048
2,048
null
-0.01713
0.022033
-0.019952
12
-0.002014
0.005002
-0.030053
-0.041454
47.387793
0.0874
0.013981
0.204946
42.801631
0.802468
0.540415
0.009486
0.412919
0.969587
0.501399
14
0.620833
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-4096
4,096
null
-0.018255
0.018208
-0.020398
12
-0.007523
-0.003222
-0.019517
-0.04276
72.876877
0.090362
0.015487
0.204253
39.399028
0.832385
0.561168
0.00477
0.418599
0.968807
0.489217
14
0.604167
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-8192
8,192
null
-0.022322
0.026785
-0.020992
12
-0.020773
0.007649
-0.016796
-0.059366
105.520056
0.103056
0.034132
0.226947
35.442318
0.828383
0.54893
0.004724
0.415855
0.96921
0.491173
14
0.677083
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-16384
16,384
null
-0.031699
0.03175
-0.022696
12
-0.039895
-0.009618
-0.01141
-0.065873
157.081765
0.1326
0.151525
0.178616
47.14418
0.790637
0.549257
0.008682
0.419301
0.968655
0.480496
14
0.697917
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-32768
32,768
null
-0.011516
0.030916
-0.016734
12
-0.039577
0.034861
-0.024186
-0.017162
214.160399
0.160554
0.298529
0.119796
62.863978
0.799985
0.53504
0.00477
0.415331
0.969252
0.487245
14
0.645833
beetle-fineweb3-eng
Beetle-FineWeb3-24B/beetle-monolingual-fineweb3-eng@step-65536
65,536
null
-0.028189
0.029561
-0.029558
12
-0.066697
-0.005069
-0.032548
-0.008442
249.251353
0.179382
0.391956
0.110067
61.759419
0.895584
0.53632
0.00392
0.416886
0.969049
0.503701
13.75
0.691667
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-0
0
null
-0.018115
0.016015
-0.015973
12
0.004008
-0.025899
-0.029927
-0.020642
24.355214
0.057374
0.005501
0.363639
21.289583
null
0.52214
0.00392
0.412523
0.969651
0.439553
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-1
1
null
-0.018115
0.016015
-0.015973
12
0.004008
-0.025899
-0.029927
-0.020642
24.355214
0.057374
0.005501
0.363639
21.289583
1
0.52214
0.00392
0.412523
0.969651
0.439553
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-2
2
null
-0.018115
0.016015
-0.015973
12
0.004008
-0.025899
-0.029927
-0.020642
24.355214
0.057374
0.005501
0.363639
21.289583
1
0.52214
0.00392
0.412523
0.969651
0.439553
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-4
4
null
-0.018115
0.016015
-0.015973
12
0.004008
-0.025899
-0.029927
-0.020642
24.355214
0.057374
0.005501
0.363639
21.289583
1
0.52214
0.00392
0.412523
0.969651
0.439553
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-6
6
null
-0.018115
0.016015
-0.015973
12
0.004008
-0.025899
-0.029927
-0.020642
24.355214
0.057374
0.005501
0.363639
21.289583
1
0.52214
0.00392
0.412523
0.969651
0.439553
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-8
8
null
-0.018115
0.016015
-0.015973
12
0.004008
-0.025899
-0.029927
-0.020642
24.355214
0.057374
0.005501
0.363639
21.289583
1
0.52214
0.00392
0.412523
0.969651
0.439553
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-16
16
null
-0.018116
0.016019
-0.015972
12
0.004021
-0.025921
-0.029895
-0.020669
24.355157
0.057373
0.0055
0.36364
21.289319
1
0.522167
0.00392
0.412523
0.969651
0.439556
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-32
32
null
-0.01814
0.015994
-0.015964
12
0.003954
-0.02587
-0.029852
-0.02079
24.354813
0.057368
0.005499
0.363646
21.287935
0.999999
0.522155
0.00392
0.412522
0.969652
0.440063
14
0.429167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-64
64
null
-0.018193
0.015972
-0.015916
12
0.003845
-0.025762
-0.029837
-0.021018
24.353855
0.057347
0.005493
0.363684
21.279042
0.999987
0.522538
0.00392
0.41252
0.969654
0.438506
14
0.452083
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-256
256
null
-0.010175
0.024329
-0.00703
12
-0.006917
-0.028479
-0.030638
0.025336
27.42184
0.086518
0.014318
0.234245
41.429783
0.777641
0.541973
0.001565
0.414488
0.969354
0.451725
14
0.4625
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-500
500
null
-0.007423
0.031397
-0.005219
12
-0.008826
-0.025382
-0.035481
0.039996
33.033007
0.109988
0.024527
0.217707
57.525921
0.951347
0.530417
0.00392
0.417454
0.968893
0.404629
14
0.54375
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-1000
1,000
null
-0.012996
0.037474
-0.009323
12
-0.000528
-0.044359
-0.044925
0.037826
33.498123
0.103468
0.022805
0.294265
34.352241
0.793099
0.540534
0.005499
0.418935
0.968705
0.418054
14
0.53125
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-2000
2,000
null
-0.020982
0.033801
-0.002283
12
0.030674
-0.051389
-0.02553
-0.037683
32.053577
0.090324
0.016753
0.305349
29.765043
0.817234
0.520322
0.00313
0.416874
0.969042
0.4697
14
0.583333
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-2500
2,500
null
-0.015234
0.029328
-0.007557
12
0.009441
-0.003412
-0.023574
-0.043393
31.704666
0.083142
0.013477
0.292932
31.12569
0.940053
0.512679
0.005552
0.414881
0.969291
0.513808
14
0.61875
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-4000
4,000
null
-0.019841
0.030877
-0.01329
12
0.016069
-0.012735
-0.024225
-0.058475
40.842382
0.088518
0.014202
0.218536
43.802865
0.816924
0.555365
0.006297
0.418354
0.9688
0.487174
14
0.579167
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-5500
5,500
null
-0.012242
0.037566
-0.014648
12
0.041397
-0.015024
-0.029297
-0.046043
52.763214
0.103985
0.020661
0.176918
51.926837
0.753171
0.539624
0.006304
0.417266
0.968947
0.510355
14
0.6625
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-7500
7,500
null
-0.003373
0.048263
-0.003526
12
0.052854
0.020376
-0.030787
-0.055937
67.420529
0.103132
0.020231
0.18291
55.129824
0.771677
0.55787
0.01112
0.417019
0.969026
0.526341
14
0.620833
beetle-humanscale-eng
Beetle-HumanScale/beetle-monolingual-humanscale-eng@step-9500
9,500
null
-0.018565
0.031301
-0.015728
12
0.021438
-0.012142
-0.031133
-0.052422
67.956088
0.099832
0.018853
0.196186
60.850795
0.85122
0.547097
0.003935
0.418896
0.968749
0.491977
14
0.685417
pico-decoder-large
pico-lm/pico-decoder-large@daa12d1eb44fa1ea4cc31aa35a8cb029917302e9
0
null
-0.012295
0.023097
-0.006458
12
-0.005136
0.014933
-0.018531
-0.040445
31.51614
0.056756
0.005229
0.391296
17.076346
null
0.54069
0.004581
0.417428
0.970631
0.472217
12
0.527083
pico-decoder-large
pico-lm/pico-decoder-large@70ace66cdec9f20ba70c65109b42b9c263e0139b
1,000
null
-0.012078
0.033119
-0.005986
12
-0.006976
0.030561
-0.017502
-0.054394
36.693862
0.079561
0.011875
0.356249
20.164182
0.932397
0.53605
0.006852
0.415889
0.970875
0.501077
12
0.595833
pico-decoder-large
pico-lm/pico-decoder-large@b7e416710019991ae8f34506fbea7a6ac31a8252
2,000
null
-0.004093
0.024135
-0.000987
12
0.003189
0.019312
-0.035809
-0.003065
293.626644
0.285179
0.128442
0.02801
227.774616
0.292421
0.462618
0.002273
0.383607
0.974941
0.461559
11.5
0.577083
pico-decoder-large
pico-lm/pico-decoder-large@d43df3f024539a5db7dfcab339e807d4842bcffd
3,000
null
-0.008209
0.028859
-0.005579
12
-0.01362
0.031273
-0.008361
-0.042127
67.734859
0.170237
0.061137
0.148067
37.5644
0.327176
0.542655
0.00273
0.418782
0.970488
0.528412
12
0.660417
pico-decoder-large
pico-lm/pico-decoder-large@23e4297bb64e81b073da327cba679d797ca4fe1d
4,000
null
-0.017893
0.018344
-0.006747
12
-0.020201
-0.022575
-0.018503
-0.010293
78.15689
0.187303
0.07966
0.184328
36.638798
0.812318
0.550202
0.006423
0.418365
0.970515
0.521063
12
0.695833
pico-decoder-large
pico-lm/pico-decoder-large@cab5aa27e02bdec8e8bf68239ae1e6758956e134
5,000
null
-0.03157
0.021197
-0.025951
12
-0.039666
-0.037931
-0.010611
-0.038071
125.586822
0.199152
0.119351
0.111392
49.983833
0.734979
0.54392
0.003643
0.414495
0.971037
0.49408
12
0.670833
pico-decoder-large
pico-lm/pico-decoder-large@e556a1a2b5a081ce269a811dfdfa329e230c674f
7,000
null
-0.017235
0.018134
-0.010625
12
-0.021187
0.003047
-0.021046
-0.029753
145.145897
0.210438
0.235561
0.148424
49.076117
0.817836
0.539491
0.006388
0.415102
0.970934
0.515539
12
0.685417
pico-decoder-large
pico-lm/pico-decoder-large@1eb202826693ab699be3043f3579cbde8211a5ef
8,000
null
-0.013727
0.03466
-0.005355
12
-0.043305
0.025347
-0.014382
-0.022569
169.140952
0.216858
0.324057
0.13124
58.175337
0.85381
0.536189
0.005474
0.417294
0.970655
0.541768
12
0.685417
pico-decoder-large
pico-lm/pico-decoder-large@94cc0caf8c3cb477f0dc32aa92a7b8e182f710bf
10,000
null
-0.001873
0.034269
0.000175
12
-0.028581
0.049398
-0.014834
-0.013475
216.43841
0.21272
0.476275
0.12265
58.378171
0.833273
0.548313
0.005477
0.416793
0.970722
0.558507
12
0.7375
pico-decoder-large
pico-lm/pico-decoder-large@10d61c2f23ed2b432c011ff4735614817452abfc
13,000
null
-0.019089
0.016941
-0.018258
12
-0.025519
-0.014942
-0.01869
-0.017206
270.517181
0.219231
0.589001
0.093158
58.850256
0.805081
0.546693
0.005025
0.418127
0.97054
0.58061
12
0.7625
pico-decoder-large
pico-lm/pico-decoder-large@609cb555fe8ab89fe8deccc19ae52176c053de3c
16,000
null
-0.011616
0.021015
-0.009881
12
0.000535
-0.002663
-0.008188
-0.03615
348.906738
0.22212
0.617699
0.098403
69.152033
0.787537
0.566531
0.006854
0.416429
0.970795
0.554802
12
0.725
pico-decoder-large
pico-lm/pico-decoder-large@2e9b09372c42b65faa9486c0d7b8191ff5b5c4de
20,000
null
-0.019792
0.017849
-0.012357
12
-0.040291
-0.014168
-0.012644
-0.012063
402.861328
0.226064
0.664062
0.105833
79.010652
0.766511
0.556002
0.008706
0.41961
0.970304
0.535027
12
0.735417
pico-decoder-large
pico-lm/pico-decoder-large@020ad4d5ea0fe4c387769324ee38ad1d15bb3f85
24,000
null
0.002562
0.020409
0.002599
12
0.022317
-0.005745
-0.007507
0.001185
471.249789
0.23799
0.70008
0.0674
93.796846
0.646754
0.55402
0.00502
0.416213
0.970808
0.549657
12
0.716667
pico-decoder-large
pico-lm/pico-decoder-large@dab022094801d1fd4c0c4deb04d7c0815ea0d9b4
30,000
null
-0.004984
0.019631
0.000084
12
-0.003518
-0.006564
-0.013246
0.003393
549.266559
0.24647
0.730602
0.055437
107.970566
0.649466
0.548304
0.006847
0.411535
0.971448
0.629679
12
0.722917
pico-decoder-large
pico-lm/pico-decoder-large@22fe7c47349efbd4d073c95fe64e3b50251cb0fb
37,000
null
-0.013347
0.025253
-0.009418
12
-0.01072
-0.035976
-0.01231
0.005618
639.421795
0.255537
0.74433
0.037307
129.937477
0.814383
0.56533
0.007321
0.415009
0.97095
0.601443
12
0.691667
pico-decoder-large
pico-lm/pico-decoder-large@5a81c54569b9cadfac8813c3db918fae371226c7
45,000
null
-0.004263
0.026304
-0.003169
12
0.012632
-0.026663
-0.023663
0.02064
666.340083
0.252668
0.761278
0.024366
154.822628
0.84422
0.53093
0.010117
0.411493
0.971416
0.625108
12
0.69375
pico-decoder-large
pico-lm/pico-decoder-large@ea63999d87b948bca4d4e9938706f51ccc88c365
55,000
null
-0.002593
0.024649
-0.000493
12
-0.005786
0.017725
-0.022409
0.000097
708.856119
0.259565
0.772065
0.021334
194.698196
0.907292
0.534561
0.006454
0.412379
0.971323
0.634063
12
0.69375
pico-decoder-large
pico-lm/pico-decoder-large@6dc96c3465c47b8b192056175e156d703a132e5e
68,000
null
-0.016651
0.02214
-0.00371
12
-0.036259
-0.009852
-0.01644
-0.004054
735.035731
0.263451
0.782708
0.017378
189.793698
0.895479
0.522462
0.013023
0.407258
0.971976
0.675709
12
0.722917
pico-decoder-large
pico-lm/pico-decoder-large@46310457a1f616890732dc9d54ca3b6f35263f47
83,000
null
-0.012794
0.019402
-0.005995
12
-0.01703
-0.01361
-0.022216
0.001682
767.563453
0.270525
0.783606
0.01674
234.649531
0.933919
0.521927
0.011499
0.404514
0.972245
0.653294
12
0.714583
pico-decoder-large
pico-lm/pico-decoder-large@1ca25b1c86203affda1776487d77204c79b5c4f7
102,000
null
-0.01427
0.016476
-0.011016
12
-0.024689
-0.003746
-0.01631
-0.012338
781.781438
0.279368
0.786632
0.015261
217.506835
0.94141
0.540209
0.008714
0.404734
0.972272
0.655684
11.5
0.725
pico-decoder-large
pico-lm/pico-decoder-large@dc35be996509151e2b2a2066a545a54a6e4900cb
125,000
null
-0.025313
0.017884
-0.01752
12
-0.040162
-0.026818
-0.014209
-0.020064
743.218853
0.281922
0.784415
0.014911
214.108679
0.954576
0.533322
0.008724
0.40222
0.972579
0.649586
11.5
0.69375
pico-decoder-medium
pico-lm/pico-decoder-medium@a6f849d1da39698d655fd3552435d28fba898dc9
0
null
0.001096
0.023415
-0.003483
12
0.007615
0.02572
-0.034605
0.005655
22.23421
0.056143
0.005265
0.366418
18.436461
null
0.54944
0.007449
0.419947
0.967954
0.463498
12
0.502083
pico-decoder-medium
pico-lm/pico-decoder-medium@b09eab97230964af5bf3964d902db96684add7f0
1,000
null
-0.004185
0.031875
-0.005004
12
-0.001077
0.041614
-0.026008
-0.03127
26.832838
0.088736
0.015673
0.319624
25.185132
0.888648
0.538469
0.007388
0.419757
0.967973
0.501166
12
0.570833
pico-decoder-medium
pico-lm/pico-decoder-medium@e8ea2825c02325ec936fec3fbf2cebf7dcd7e5fd
2,000
null
0.005663
0.0247
0.005024
12
0.037124
0.00588
-0.02411
0.003757
24.588758
0.088354
0.015659
0.33519
24.278636
0.912209
0.53405
0.006446
0.418299
0.968274
0.482875
12
0.610417
pico-decoder-medium
pico-lm/pico-decoder-medium@7d53111bd8fb071fbb1f180e4435cfea1b3e3149
3,000
null
-0.006365
0.021274
-0.003813
12
0.005929
0.01347
-0.020769
-0.02409
28.475886
0.113611
0.030331
0.325063
26.091651
0.911356
0.544537
0.003704
0.414754
0.968792
0.511745
12
0.660417
pico-decoder-medium
pico-lm/pico-decoder-medium@94db452b6d8323e895d44d7e57364a44ce9d64b9
4,000
null
-0.019528
0.025391
-0.011755
12
0.00007
-0.026523
-0.014732
-0.036927
34.748915
0.136989
0.045825
0.298473
30.135428
0.899632
0.56192
0.006415
0.418398
0.968252
0.528104
12
0.704167
pico-decoder-medium
pico-lm/pico-decoder-medium@2c8ff3af5ea6a6720939283915670af5878a5fad
5,000
null
-0.011614
0.03164
0.009637
12
0.008661
0.014405
-0.016302
-0.053221
503.33167
0.360724
0.200008
0.010186
773.752983
0.150488
0.431997
0.007346
0.361665
0.975748
0.343767
11.75
0.610417
End of preview. Expand in Data Studio

CDL DevAI results — brain × interpretability × localisation, per model per checkpoint

Developmental analysis of 10 language-model families against the ds003604 auditory language fMRI dataset. For every training checkpoint of every model we measured three things and here report them side by side:

axis what it asks source tables
brain does the model's representational geometry match the brain's? brain_alignment
interp how is the representation organised internally? interp_mechanistic, interp_layerwise
localisation are linguistic phenomena isolated into dedicated units? localisation_isolation, localisation_onset

Start with the summary_by_checkpoint config (the default, and what the viewer shows first): one row per model × checkpoint, with all three axes as columns. 262 rows, 10 models.


⚠️ READ THIS BEFORE USING THE BRAIN ALIGNMENT NUMBERS

The brain alignment columns are confounded by scanner run and must not be read as a result about language models. This affects every rsa* column in brain_alignment, every brain_* column in the summary tables, and the ablation_alignment table.

In ds003604 each stimulus is presented in exactly one scanner run (for Phon, run-01 carries 48 of the 96 stimuli and run-02 the other 48). Run membership is therefore perfectly confounded with stimulus identity, and every cross-run stimulus pair inherits that run's drift, baseline shift and scaling. Measured:

"different run" predicts brain dissimilarity, Spearman rho, all 12 task × session cells:
    Gram   +0.866  +0.812  +0.828        Plaus  +0.741  +0.689  +0.761
    Phon   +0.562  +0.488  +0.624        Sem    +0.511  +0.510  +0.601

By comparison no stimulus property predicts these RDMs at all: trial type ≈ −0.02, text length ≈ 0, lexical overlap ≈ 0. The RDMs look highly reliable across independent subject cohorts (rho 0.74–0.92 between sessions of the same task) — but that is largely the reliability of an acquisition artefact, because run assignment is fixed by the protocol and so repeats identically for every subject.

A language model cannot represent which scanner run a stimulus appeared in, so its RSA against this structure is ≈0 by construction, and slightly negative in practice.

Two obvious explanations were tested and ruled out:

  • Cohort size. Sem/ses-7 was built from 40 subjects; rebuilt from 98 it agrees with the 40-subject version at rho = 0.928, identical stimuli. Bigger cohorts change nothing.
  • Layer choice. Alignment is flat across every layer — −0.012 to −0.021 across all 12 layers of babylm-gpt2-3 and all 14 of beetle-humanscale-eng. There is no middle-layer peak being missed. See diagnostics_layerwise.

What fixes it, and what happens when you do. Z-scoring each voxel within run before combining runs removes the confound cleanly — run predictiveness falls from +0.562 to −0.041. Alignment against the corrected RDM then peaks at +0.022 (layer 2, beetle-humanscale-eng, 2,556 stimulus pairs), which is not significant. The layer profile becomes sensible — early/middle layers positive, output layers negative — but the magnitude is ≈0. See diagnostics_run_confound.

Bottom line: these tables do not show that language models align with the brain, and they do not show that they fail to. The measurement cannot answer it. The corrected analysis, on one cell and one model, finds no detectable alignment. Do not cite the raw numbers in either direction.

Which columns are safe

axis affected by the run confound?
brain_*, rsa*, ablation_alignment Confounded. Do not use as a model result.
interp_* (norm, gini, hoyer, per, condition_number, cka_to_prev) ✅ Safe — computed from LM activations only. The fMRI data is not involved.
loc_* / localisation_* (selectivity_index, overlap, gini, entropy, layer_com, n_active_layers) ✅ Safe — LM-internal localisation against text contrasts. No fMRI.
behaviour (mp_accuracy) ✅ Safe — minimal-pair accuracy, text only.
ablation_behaviour (causal_selectivity) ✅ Safe — ablation vs behaviour, no fMRI.

The one uncontaminated positive result in this release is the causal behaviour test: ablating a phenomenon's localized circuit costs 1.13% minimal-pair accuracy versus 0.55% for a random circuit of the same size — selectivity +0.0058, t = 1.98, p = 0.049, n = 316, driven mostly by Phon (+0.021). That is borderline and should be described as suggestive, not established.


Layout

overall/
  by_checkpoint.csv        <- THE MAIN TABLE. one row per (family, step),
                              brain + interp + localisation side by side
  summary_by_family.csv    <- one row per model: means, ranges, step-vs-alignment trend
  claim_tests.csv          <- per-family claim tests (claim, stat, value, p, n)
  heldout_predictor.csv    <- cross-family held-out predictive validation
  localisation_onset.csv
by-model/<family>/
  README.md                <- what this model is, its numbers, what is odd about it
  checkpoints.csv          <- this model's rows of the main table
  brain_alignment.csv      <- full per task × session × step detail  (CONFOUNDED)
  interp_mechanistic.csv   interp_layerwise.csv
  localisation_isolation.csv  localisation_onset.csv
  behaviour.csv            ablation_alignment.csv  ablation_behaviour.csv
  figures/<family>_overview.png
diagnostics/
  layerwise_alignment.csv  <- alignment at every layer (rules out the layer explanation)
  run_confound_check.csv   <- raw vs run-partialled alignment, per layer
figures/                   <- cross-model figures (fig1-fig8, tables)
superseded/early_tier1/    <- an earlier PARTIAL pass (7 families, 8/12 cells). Kept for
                              completeness, deliberately NOT a viewer config. Do not use.
provenance_tier_ledger.json <- the run record: per-tier status, exit code, duration, peak GPU

Every table carries family and model_ref, so you can filter by model in the viewer without downloading anything.


The models

10 families, 262 checkpoints total. brain_rsa_mean is shown only so you can see it is flat and near zero; per the warning above it is not interpretable.

family ckpts steps brain RSA (confounded) trend ρ (p) interp PR interp gini loc selectivity behaviour acc
pico-decoder-tiny 21 0–126k −0.017 −0.02 (0.80) 0.221 0.204 0.546 0.631
pico-decoder-small 126 0–125k +0.001 +0.16 (<0.001) 0.184 0.243 0.551 0.682
pico-decoder-medium 21 0–125k −0.011 −0.10 (0.13) 0.191 0.206 0.544 0.679
pico-decoder-large 21 0–125k −0.012 +0.01 (0.93) 0.104 0.221 0.539 0.687
beetle-humanscale-eng 18 0–9.5k −0.016 −0.03 (0.71) 0.300 0.077 0.530 0.510
beetle-fineweb3-eng 19 0–65k −0.015 −0.05 (0.44) 0.241 0.090 0.558 0.590
babylm-gpt2-3 9 191–1908 −0.021 +0.17 (0.07) 0.065 0.326 0.519 0.693
babylm-gpt2-5 9 191–1908 −0.022 +0.14 (0.14) 0.066 0.324 0.531 0.687
babylm-gpt2-7 9 191–1908 −0.024 +0.12 (0.21) 0.065 0.328 0.520 0.707
babylm-gpt2 9 598–2990 −0.032 −0.03 (0.73) 0.072 0.293 0.530 0.707

pico-decoder-small's trend is p<0.001 only because n = 1,512 stimulus-level rows; ρ = 0.16 on a confounded measure is not a finding. No other family reaches p < 0.05, which across ten tests is what chance looks like. The held-out cross-family predictor scores mean R² = −2.74 — worse than predicting the mean.


How to read the metrics

interp (safe). per = participation ratio, the effective dimensionality of the representation as a fraction of hidden size; lower = more compressed. Here 0.06–0.30, and the babylm models (0.065) are far more compressed than Beetle (0.24–0.30). gini and hoyer are sparsity of activation mass, higher = sparser. condition_number is the spread of the activation covariance spectrum. cka_to_prev is representational similarity to the previous checkpoint — near 1 means training has stopped changing the geometry.

localisation (safe). selectivity_index is how strongly a phenomenon's top units prefer it over other phenomena; ~0.5 across all models here, i.e. moderate and strikingly constant — no model isolates phenomena sharply. mean_overlap_with_others is how much a phenomenon's circuit is shared with other phenomena; high overlap means little specialisation. n_active_layers is how many layers contribute; layer_com is the centre of mass over depth (low = early layers).

behaviour (safe). mp_accuracy is minimal-pair accuracy, chance = 0.5. Beetle humanscale at 0.51 is essentially at chance; the babylm and pico models reach 0.63–0.71.

brain (confounded — see the warning). rsa is Spearman between the LM RDM and the brain RDM over stimulus pairs; rsa_pearson/rsa_kendall are the same with different rank treatments. n_stim is 72 for Sem/Phon and 60 for Gram/Plaus (controls excluded).


Provenance

Produced by suchirsalhan/cdl-representations-brains-babylms, tiers 0–3, 2026-08-19. Brain-side session RDMs are cached separately at BrainAlign/ds003604-session-rdms (12 of 12 task × session cells). The run confound described above applies to those RDMs too, and the recommended fix is to normalise voxel patterns within run before aggregating across runs.

Downloads last month
44