geolip-vit-x34 / expert_analysis_output.txt
AbstractPhil's picture
Update expert_analysis_output.txt
13a294a verified
=================================================================
34-EXPERT BOUNDARY SCAN
=================================================================
Loaded 34 experts, N=5000
=================================================================
SCAN 1: PER-EXPERT DISTRIBUTIONS
=================================================================
Expert dim norm_μ norm_σ eff_dim top1_sv sparsity l2_normed
-----------------------------------------------------------------------------------------------
clip_b16_laion2b 512 12.786 0.5695 277.9 0.0140 0.000 ✗
clip_b16_openai 512 10.744 0.5222 275.8 0.0146 0.000 ✗
clip_b32_datacomp 512 14.834 0.6564 282.3 0.0145 0.000 ✗
clip_b32_laion2b 512 10.710 0.5407 284.0 0.0143 0.000 ✗
clip_b32_openai 512 10.580 0.5588 279.2 0.0145 0.000 ✗
clip_bigg14_laion2b 1280 41.838 1.4545 391.7 0.0098 0.000 ✗
clip_g14_laion2b 1024 27.776 1.2458 360.4 0.0102 0.000 ✗
clip_h14_laion2b 1024 20.983 0.8719 374.3 0.0103 0.000 ✗
clip_l14_336_openai 768 18.795 0.8237 370.8 0.0113 0.000 ✗
clip_l14_datacomp 768 21.482 0.8630 349.7 0.0110 0.000 ✗
clip_l14_laion2b 768 19.414 0.8304 317.5 0.0124 0.000 ✗
clip_l14_openai 768 19.120 0.8477 371.3 0.0112 0.000 ✗
dinov2_b14 768 46.439 0.9252 524.6 0.0078 0.000 ✗
dinov2_b14_reg 768 24.345 0.8091 505.4 0.0079 0.000 ✗
dinov2_g14 1536 52.150 1.6881 698.8 0.0067 0.000 ✗
dinov2_g14_reg 1536 49.784 2.6030 730.2 0.0065 0.000 ✗
dinov2_l14 1024 45.182 0.6897 667.1 0.0062 0.000 ✗
dinov2_l14_reg 1024 22.890 0.7109 635.4 0.0063 0.001 ✗
dinov2_s14 384 47.601 1.3296 293.5 0.0101 0.000 ✗
dinov2_s14_reg 384 25.019 0.9930 289.7 0.0103 0.000 ✗
mae_b16 768 16.640 0.9910 33.9 0.1479 0.001 ✗
mae_h14 1280 21.486 0.3228 240.6 0.0249 0.001 ✗
mae_l16 1024 16.649 0.1999 206.6 0.0333 0.001 ✗
siglip2_b16_256 768 11.186 0.6884 335.0 0.0129 0.000 ✗
siglip2_b16_512 768 10.870 0.7466 338.1 0.0126 0.000 ✗
siglip2_l16_384 1024 15.719 1.1567 369.6 0.0116 0.000 ✗
siglip_b16_384 768 14.471 1.0706 356.0 0.0130 0.000 ✗
siglip_b16_512 768 14.304 1.0488 356.2 0.0128 0.000 ✗
siglip_l16_256 1024 26.095 1.6823 357.2 0.0127 0.000 ✗
siglip_l16_384 1024 24.232 1.6070 361.5 0.0128 0.000 ✗
siglip_so400m_384 1152 19.771 1.4144 409.2 0.0115 0.000 ✗
vit_b16_21k 768 89.054 4.2170 527.9 0.0089 0.000 ✗
vit_l16_21k 1024 85.693 3.4700 673.3 0.0075 0.004 ✗
vit_s16_21k 384 46.640 2.8304 280.8 0.0139 0.000 ✗
=================================================================
SCAN 2: PAIRWISE RAW COSINE (same-dim experts only)
=================================================================
384-d (3 experts):
dinov2_s14 × dinov2_s14_reg cos=0.0049 ▼
dinov2_s14 × vit_s16_21k cos=0.0018 ▼
dinov2_s14_reg × vit_s16_21k cos=-0.0017 ▼
512-d (5 experts):
clip_b16_laion2b × clip_b16_openai cos=-0.0076 ▼
clip_b16_laion2b × clip_b32_datacomp cos=-0.0186 ▼
clip_b16_laion2b × clip_b32_laion2b cos=-0.0129 ▼
clip_b16_laion2b × clip_b32_openai cos=0.0104 ▼
clip_b16_openai × clip_b32_datacomp cos=-0.0134 ▼
clip_b16_openai × clip_b32_laion2b cos=-0.0041 ▼
clip_b16_openai × clip_b32_openai cos=0.0096 ▼
clip_b32_datacomp × clip_b32_laion2b cos=0.0353 ▼
clip_b32_datacomp × clip_b32_openai cos=0.0062 ▼
clip_b32_laion2b × clip_b32_openai cos=0.0097 ▼
768-d (12 experts):
clip_l14_336_openai × clip_l14_datacomp cos=-0.0111 ▼
clip_l14_336_openai × clip_l14_laion2b cos=-0.0087 ▼
clip_l14_336_openai × clip_l14_openai cos=0.9585 ★
clip_l14_336_openai × dinov2_b14 cos=0.0011 ▼
clip_l14_336_openai × dinov2_b14_reg cos=0.0046 ▼
clip_l14_336_openai × mae_b16 cos=0.0095 ▼
clip_l14_336_openai × siglip2_b16_256 cos=0.0026 ▼
clip_l14_336_openai × siglip2_b16_512 cos=0.0023 ▼
clip_l14_336_openai × siglip_b16_384 cos=0.0111 ▼
clip_l14_336_openai × siglip_b16_512 cos=0.0109 ▼
clip_l14_336_openai × vit_b16_21k cos=-0.0025 ▼
clip_l14_datacomp × clip_l14_laion2b cos=-0.0052 ▼
clip_l14_datacomp × clip_l14_openai cos=-0.0079 ▼
clip_l14_datacomp × dinov2_b14 cos=-0.0024 ▼
clip_l14_datacomp × dinov2_b14_reg cos=0.0060 ▼
clip_l14_datacomp × mae_b16 cos=0.0123 ▼
clip_l14_datacomp × siglip2_b16_256 cos=0.0118 ▼
clip_l14_datacomp × siglip2_b16_512 cos=0.0135 ▼
clip_l14_datacomp × siglip_b16_384 cos=-0.0092 ▼
clip_l14_datacomp × siglip_b16_512 cos=-0.0105 ▼
clip_l14_datacomp × vit_b16_21k cos=0.0071 ▼
clip_l14_laion2b × clip_l14_openai cos=-0.0076 ▼
clip_l14_laion2b × dinov2_b14 cos=0.0043 ▼
clip_l14_laion2b × dinov2_b14_reg cos=-0.0006 ▼
clip_l14_laion2b × mae_b16 cos=-0.0055 ▼
clip_l14_laion2b × siglip2_b16_256 cos=-0.0200 ▼
clip_l14_laion2b × siglip2_b16_512 cos=-0.0224 ▼
clip_l14_laion2b × siglip_b16_384 cos=0.0076 ▼
clip_l14_laion2b × siglip_b16_512 cos=0.0070 ▼
clip_l14_laion2b × vit_b16_21k cos=-0.0075 ▼
clip_l14_openai × dinov2_b14 cos=0.0000 ▼
clip_l14_openai × dinov2_b14_reg cos=0.0045 ▼
clip_l14_openai × mae_b16 cos=0.0085 ▼
clip_l14_openai × siglip2_b16_256 cos=0.0021 ▼
clip_l14_openai × siglip2_b16_512 cos=0.0004 ▼
clip_l14_openai × siglip_b16_384 cos=0.0096 ▼
clip_l14_openai × siglip_b16_512 cos=0.0096 ▼
clip_l14_openai × vit_b16_21k cos=-0.0009 ▼
dinov2_b14 × dinov2_b14_reg cos=-0.0044 ▼
dinov2_b14 × mae_b16 cos=0.0028 ▼
dinov2_b14 × siglip2_b16_256 cos=0.0025 ▼
dinov2_b14 × siglip2_b16_512 cos=0.0029 ▼
dinov2_b14 × siglip_b16_384 cos=0.0012 ▼
dinov2_b14 × siglip_b16_512 cos=0.0008 ▼
dinov2_b14 × vit_b16_21k cos=0.0028 ▼
dinov2_b14_reg × mae_b16 cos=0.0036 ▼
dinov2_b14_reg × siglip2_b16_256 cos=0.0000 ▼
dinov2_b14_reg × siglip2_b16_512 cos=-0.0009 ▼
dinov2_b14_reg × siglip_b16_384 cos=0.0019 ▼
dinov2_b14_reg × siglip_b16_512 cos=0.0012 ▼
dinov2_b14_reg × vit_b16_21k cos=-0.0008 ▼
mae_b16 × siglip2_b16_256 cos=-0.0036 ▼
mae_b16 × siglip2_b16_512 cos=-0.0013 ▼
mae_b16 × siglip_b16_384 cos=0.0053 ▼
mae_b16 × siglip_b16_512 cos=0.0048 ▼
mae_b16 × vit_b16_21k cos=-0.0043 ▼
siglip2_b16_256 × siglip2_b16_512 cos=0.9545 ★
siglip2_b16_256 × siglip_b16_384 cos=-0.0051 ▼
siglip2_b16_256 × siglip_b16_512 cos=-0.0070 ▼
siglip2_b16_256 × vit_b16_21k cos=0.0011 ▼
siglip2_b16_512 × siglip_b16_384 cos=-0.0038 ▼
siglip2_b16_512 × siglip_b16_512 cos=-0.0052 ▼
siglip2_b16_512 × vit_b16_21k cos=0.0018 ▼
siglip_b16_384 × siglip_b16_512 cos=0.9741 ★
siglip_b16_384 × vit_b16_21k cos=-0.0057 ▼
siglip_b16_512 × vit_b16_21k cos=-0.0057 ▼
1024-d (9 experts):
clip_g14_laion2b × clip_h14_laion2b cos=-0.0185 ▼
clip_g14_laion2b × dinov2_l14 cos=-0.0036 ▼
clip_g14_laion2b × dinov2_l14_reg cos=0.0011 ▼
clip_g14_laion2b × mae_l16 cos=-0.0031 ▼
clip_g14_laion2b × siglip2_l16_384 cos=-0.0033 ▼
clip_g14_laion2b × siglip_l16_256 cos=0.0051 ▼
clip_g14_laion2b × siglip_l16_384 cos=0.0057 ▼
clip_g14_laion2b × vit_l16_21k cos=0.0015 ▼
clip_h14_laion2b × dinov2_l14 cos=-0.0028 ▼
clip_h14_laion2b × dinov2_l14_reg cos=-0.0053 ▼
clip_h14_laion2b × mae_l16 cos=-0.0016 ▼
clip_h14_laion2b × siglip2_l16_384 cos=-0.0136 ▼
clip_h14_laion2b × siglip_l16_256 cos=-0.0146 ▼
clip_h14_laion2b × siglip_l16_384 cos=-0.0130 ▼
clip_h14_laion2b × vit_l16_21k cos=-0.0011 ▼
dinov2_l14 × dinov2_l14_reg cos=0.0008 ▼
dinov2_l14 × mae_l16 cos=-0.0019 ▼
dinov2_l14 × siglip2_l16_384 cos=0.0002 ▼
dinov2_l14 × siglip_l16_256 cos=0.0022 ▼
dinov2_l14 × siglip_l16_384 cos=0.0016 ▼
dinov2_l14 × vit_l16_21k cos=0.0009 ▼
dinov2_l14_reg × mae_l16 cos=0.0054 ▼
dinov2_l14_reg × siglip2_l16_384 cos=0.0054 ▼
dinov2_l14_reg × siglip_l16_256 cos=-0.0023 ▼
dinov2_l14_reg × siglip_l16_384 cos=-0.0019 ▼
dinov2_l14_reg × vit_l16_21k cos=-0.0015 ▼
mae_l16 × siglip2_l16_384 cos=0.0206 ▼
mae_l16 × siglip_l16_256 cos=0.0035 ▼
mae_l16 × siglip_l16_384 cos=0.0021 ▼
mae_l16 × vit_l16_21k cos=0.0135 ▼
siglip2_l16_384 × siglip_l16_256 cos=-0.0057 ▼
siglip2_l16_384 × siglip_l16_384 cos=-0.0054 ▼
siglip2_l16_384 × vit_l16_21k cos=-0.0109 ▼
siglip_l16_256 × siglip_l16_384 cos=0.9535 ★
siglip_l16_256 × vit_l16_21k cos=0.0009 ▼
siglip_l16_384 × vit_l16_21k cos=0.0001 ▼
1280-d (2 experts):
clip_bigg14_laion2b × mae_h14 cos=0.0049 ▼
1536-d (2 experts):
dinov2_g14 × dinov2_g14_reg cos=-0.0000 ▼
=================================================================
SCAN 3: PROCRUSTES ALIGNMENT (all pairs, projected to min dim)
=================================================================
Pairwise Procrustes alignment (19 representatives):
clip_b clip_b clip_l clip_g clip_h clip_b dinov2 dinov2 dinov2 dinov2 mae_b1 mae_l1 mae_h1 siglip siglip siglip vit_s1 vit_b1 vit_l1
clip_b16 --- 0.618 0.614 0.598 0.607 0.604 0.499 0.530 0.527 0.544 0.496 0.507 0.505 0.579 0.578 0.583 0.501 0.533 0.534
clip_b32 0.618 --- 0.584 0.590 0.598 0.595 0.503 0.530 0.525 0.542 0.502 0.513 0.512 0.570 0.565 0.568 0.512 0.540 0.540
clip_l14 0.614 0.584 --- 0.631 0.635 0.636 0.492 0.596 0.596 0.605 0.575 0.582 0.582 0.618 0.624 0.630 0.491 0.593 0.596
clip_g14 0.598 0.590 0.631 --- 0.709 0.712 0.503 0.598 0.671 0.675 0.580 0.666 0.666 0.630 0.692 0.695 0.504 0.599 0.674
clip_h14 0.607 0.598 0.635 0.709 --- 0.713 0.511 0.601 0.674 0.679 0.580 0.665 0.666 0.637 0.695 0.698 0.509 0.602 0.675
clip_bigg14 0.604 0.595 0.636 0.712 0.713 --- 0.511 0.601 0.675 0.751 0.581 0.667 0.745 0.636 0.698 0.732 0.510 0.603 0.677
dinov2_s14 0.499 0.503 0.492 0.503 0.511 0.511 --- 0.585 0.550 0.583 0.482 0.499 0.496 0.516 0.508 0.506 0.511 0.511 0.507
dinov2_b14 0.530 0.530 0.596 0.598 0.601 0.601 0.585 --- 0.646 0.671 0.580 0.590 0.590 0.602 0.601 0.600 0.505 0.599 0.601
dinov2_l14 0.527 0.525 0.596 0.671 0.674 0.675 0.550 0.646 --- 0.726 0.576 0.664 0.664 0.602 0.675 0.675 0.495 0.597 0.672
dinov2_g14 0.544 0.542 0.605 0.675 0.679 0.751 0.583 0.671 0.726 --- 0.578 0.664 0.744 0.613 0.682 0.717 0.515 0.606 0.676
mae_b16 0.496 0.502 0.575 0.580 0.580 0.581 0.482 0.580 0.576 0.578 --- 0.636 0.630 0.578 0.577 0.576 0.480 0.591 0.593
mae_l16 0.507 0.513 0.582 0.666 0.665 0.667 0.499 0.590 0.664 0.664 0.636 --- 0.706 0.586 0.663 0.663 0.492 0.598 0.672
mae_h14 0.505 0.512 0.582 0.666 0.666 0.745 0.496 0.590 0.664 0.744 0.630 0.706 --- 0.585 0.663 0.703 0.492 0.598 0.673
siglip_b16 0.579 0.570 0.618 0.630 0.637 0.636 0.516 0.602 0.602 0.613 0.578 0.586 0.585 --- 0.663 0.653 0.508 0.599 0.602
siglip_l16 0.578 0.565 0.624 0.692 0.695 0.698 0.508 0.601 0.675 0.682 0.577 0.663 0.663 0.663 --- 0.721 0.500 0.597 0.672
siglip_so400m 0.583 0.568 0.630 0.695 0.698 0.732 0.506 0.600 0.675 0.717 0.576 0.663 0.703 0.653 0.721 --- 0.499 0.596 0.672
vit_s16 0.501 0.512 0.491 0.504 0.509 0.510 0.511 0.505 0.495 0.515 0.480 0.492 0.492 0.508 0.500 0.499 --- 0.571 0.549
vit_b16 0.533 0.540 0.593 0.599 0.602 0.603 0.511 0.599 0.597 0.606 0.591 0.598 0.598 0.599 0.597 0.596 0.571 --- 0.651
vit_l16 0.534 0.540 0.596 0.674 0.675 0.677 0.507 0.601 0.672 0.676 0.593 0.672 0.673 0.602 0.672 0.672 0.549 0.651 ---
=================================================================
SCAN 4: FAMILY CLUSTERING
=================================================================
Within-family mean cosine (raw, same-dim pairs):
clip (12 members, 17 pairs): mean=0.0538 min=-0.0186 max=0.9585
dinov2 ( 8 members, 4 pairs): mean=0.0003 min=-0.0044 max=0.0049
siglip ( 5 members, 2 pairs): mean=0.9638 min=0.9535 max=0.9741
siglip2 ( 3 members, 1 pairs): mean=0.9545 min=0.9545 max=0.9545
Patch family analysis:
patch14 (16 experts): dims=[384, 768, 1024, 1280, 1536]
clip_bigg14_laion2b dim=1280
clip_g14_laion2b dim=1024
clip_h14_laion2b dim=1024
clip_l14_336_openai dim=768
clip_l14_datacomp dim=768
clip_l14_laion2b dim=768
clip_l14_openai dim=768
dinov2_b14 dim=768
dinov2_b14_reg dim=768
dinov2_g14 dim=1536
dinov2_g14_reg dim=1536
dinov2_l14 dim=1024
dinov2_l14_reg dim=1024
dinov2_s14 dim=384
dinov2_s14_reg dim=384
mae_h14 dim=1280
patch16 (17 experts): dims=[384, 512, 768, 1024]
clip_b16_laion2b dim=512
clip_b16_openai dim=512
clip_b32_datacomp dim=512
clip_b32_laion2b dim=512
clip_b32_openai dim=512
mae_b16 dim=768
mae_l16 dim=1024
siglip2_b16_256 dim=768
siglip2_b16_512 dim=768
siglip2_l16_384 dim=1024
siglip_b16_384 dim=768
siglip_b16_512 dim=768
siglip_l16_256 dim=1024
siglip_l16_384 dim=1024
vit_b16_21k dim=768
vit_l16_21k dim=1024
vit_s16_21k dim=384
so400m ( 1 experts): dims=[1152]
siglip_so400m_384 dim=1152
=================================================================
SCAN 5: SVD SPECTRUM (effective dimensionality)
=================================================================
Expert dim eff_d ratio sv1% sv5% sv10% sv50% tail
-------------------------------------------------------------------------------------
clip_b16_laion2b 512 277.9 0.543 5.5% 19.1% 28.9% 60.2% 0.14%
clip_b16_openai 512 275.8 0.539 5.9% 21.1% 32.3% 62.9% 0.25%
clip_b32_datacomp 512 282.3 0.551 6.0% 20.3% 30.3% 60.4% 0.22%
clip_b32_laion2b 512 284.0 0.555 5.8% 19.2% 28.9% 59.8% 0.24%
clip_b32_openai 512 279.2 0.545 5.9% 21.0% 31.8% 62.2% 0.26%
clip_bigg14_laion2b 1280 391.7 0.306 3.7% 14.7% 22.8% 51.9% 0.00%
clip_g14_laion2b 1024 360.4 0.352 3.8% 15.6% 23.9% 53.3% 0.00%
clip_h14_laion2b 1024 374.3 0.366 4.0% 15.8% 24.0% 53.3% 0.00%
clip_l14_336_openai 768 370.8 0.483 4.7% 17.5% 25.9% 54.5% 0.00%
clip_l14_datacomp 768 349.7 0.455 4.2% 15.6% 24.3% 53.5% 0.00%
clip_l14_laion2b 768 317.5 0.413 4.9% 17.4% 26.2% 56.8% 0.00%
clip_l14_openai 768 371.3 0.483 4.7% 17.3% 26.0% 54.5% 0.00%
dinov2_b14 768 524.6 0.683 3.2% 10.0% 16.1% 39.2% 0.68%
dinov2_b14_reg 768 505.4 0.658 3.1% 10.6% 17.0% 41.2% 0.55%
dinov2_g14 1536 698.8 0.455 3.1% 12.1% 19.2% 43.5% 0.06%
dinov2_g14_reg 1536 730.1 0.475 3.0% 11.0% 17.7% 41.0% 0.07%
dinov2_l14 1024 667.1 0.651 2.5% 9.4% 15.0% 35.7% 0.47%
dinov2_l14_reg 1024 635.5 0.621 2.5% 9.7% 15.9% 38.0% 0.39%
dinov2_s14 384 293.5 0.764 3.0% 10.9% 17.8% 46.8% 1.15%
dinov2_s14_reg 384 289.8 0.755 3.1% 11.2% 18.4% 47.6% 1.04%
mae_b16 768 33.9 0.044 74.2% 92.3% 94.4% 98.0% 0.00%
mae_h14 1280 240.6 0.188 14.9% 47.0% 59.2% 82.6% 0.02%
mae_l16 1024 206.6 0.202 22.9% 48.3% 59.9% 83.4% 0.03%
siglip2_b16_256 768 335.0 0.436 5.6% 18.9% 28.7% 59.7% 0.02%
siglip2_b16_512 768 338.1 0.440 5.4% 18.7% 28.6% 59.5% 0.02%
siglip2_l16_384 1024 369.6 0.361 5.0% 17.1% 26.6% 56.5% 0.00%
siglip_b16_384 768 356.0 0.464 6.0% 20.0% 29.6% 59.3% 0.06%
siglip_b16_512 768 356.2 0.464 5.9% 20.0% 29.6% 59.4% 0.06%
siglip_l16_256 1024 357.2 0.349 5.8% 19.0% 28.5% 58.5% 0.00%
siglip_l16_384 1024 361.4 0.353 5.9% 19.1% 28.6% 58.0% 0.00%
siglip_so400m_384 1152 409.2 0.355 5.4% 18.2% 27.4% 55.6% 0.00%
vit_b16_21k 768 527.9 0.687 4.2% 16.2% 22.1% 42.0% 0.97%
vit_l16_21k 1024 673.3 0.657 3.8% 14.4% 20.0% 37.6% 0.57%
vit_s16_21k 384 280.8 0.731 5.4% 18.5% 26.2% 52.2% 1.49%
=================================================================
SCAN 6: OUTLIER DETECTION
=================================================================
Mean cosine to all other experts (projected to 384-d):
siglip_l16_256 1024d mean_cos=0.0189 â—„ CORE
clip_b32_datacomp 512d mean_cos=0.0095 â—„ CORE
siglip_b16_384 768d mean_cos=0.0088 â—„ CORE
clip_l14_laion2b 768d mean_cos=0.0083 â—„ CORE
dinov2_l14 1024d mean_cos=0.0082 â—„ CORE
dinov2_b14 768d mean_cos=0.0081 â—„ CORE
dinov2_g14 1536d mean_cos=0.0080 â—„ CORE
dinov2_g14_reg 1536d mean_cos=0.0079 â—„ CORE
dinov2_b14_reg 768d mean_cos=0.0064
vit_l16_21k 1024d mean_cos=0.0054
clip_h14_laion2b 1024d mean_cos=0.0052
clip_b32_openai 512d mean_cos=0.0025
siglip_b16_512 768d mean_cos=0.0024
clip_b16_openai 512d mean_cos=0.0022
clip_b32_laion2b 512d mean_cos=0.0016
dinov2_s14_reg 384d mean_cos=0.0008
mae_h14 1280d mean_cos=0.0004
vit_s16_21k 384d mean_cos=0.0004
clip_bigg14_laion2b 1280d mean_cos=-0.0005
dinov2_s14 384d mean_cos=-0.0009
mae_b16 768d mean_cos=-0.0009
clip_g14_laion2b 1024d mean_cos=-0.0017
dinov2_l14_reg 1024d mean_cos=-0.0021
clip_l14_datacomp 768d mean_cos=-0.0028
siglip_l16_384 1024d mean_cos=-0.0036
vit_b16_21k 768d mean_cos=-0.0038
siglip2_l16_384 1024d mean_cos=-0.0041
clip_l14_336_openai 768d mean_cos=-0.0069
mae_l16 1024d mean_cos=-0.0091
siglip2_b16_512 768d mean_cos=-0.0127 â—„ OUTLIER
clip_l14_openai 768d mean_cos=-0.0132 â—„ OUTLIER
clip_b16_laion2b 512d mean_cos=-0.0166 â—„ OUTLIER
siglip2_b16_256 768d mean_cos=-0.0208 â—„ OUTLIER
siglip_so400m_384 1152d mean_cos=-0.0262 â—„ OUTLIER
=================================================================
SCAN 7: VARIANCE RETAINED UNDER PROJECTION
=================================================================
Expert orig →256 →384 →512 →768 →1024
---------------------------------------------------------------------------
clip_b16_laion2b 512 0.9433 0.9891 1.000 1.000 1.000
clip_b16_openai 512 0.9379 0.9865 1.000 1.000 1.000
clip_b32_datacomp 512 0.9377 0.9875 1.000 1.000 1.000
clip_b32_laion2b 512 0.9396 0.9875 1.000 1.000 1.000
clip_b32_openai 512 0.9363 0.9861 1.000 1.000 1.000
clip_bigg14_laion2b 1280 0.8870 0.9564 0.9873 1.0000 1.0000
clip_g14_laion2b 1024 0.8988 0.9657 0.9935 1.0000 1.000
clip_h14_laion2b 1024 0.8928 0.9597 0.9893 1.0000 1.000
clip_l14_336_openai 768 0.8879 0.9564 0.9888 1.000 1.000
clip_l14_datacomp 768 0.9026 0.9689 0.9957 1.000 1.000
clip_l14_laion2b 768 0.9219 0.9786 0.9987 1.000 1.000
clip_l14_openai 768 0.8873 0.9564 0.9889 1.000 1.000
dinov2_b14 768 0.7912 0.8942 0.9533 1.000 1.000
dinov2_b14_reg 768 0.8087 0.9061 0.9599 1.000 1.000
dinov2_g14 1536 0.7817 0.8664 0.9179 0.9705 0.9912
dinov2_g14_reg 1536 0.7642 0.8546 0.9101 0.9674 0.9901
dinov2_l14 1024 0.7342 0.8418 0.9097 0.9787 1.000
dinov2_l14_reg 1024 0.7606 0.8618 0.9230 0.9824 1.000
dinov2_s14 384 0.9295 1.000 1.000 1.000 1.000
dinov2_s14_reg 384 0.9332 1.000 1.000 1.000 1.000
mae_b16 768 0.9974 0.9990 0.9997 1.000 1.000
mae_h14 1280 0.9593 0.9783 0.9880 0.9966 0.9993
mae_l16 1024 0.9676 0.9843 0.9924 0.9986 1.000
siglip2_b16_256 768 0.9221 0.9732 0.9927 1.000 1.000
siglip2_b16_512 768 0.9198 0.9716 0.9923 1.000 1.000
siglip2_l16_384 1024 0.9077 0.9660 0.9902 0.9997 1.000
siglip_b16_384 768 0.9074 0.9629 0.9878 1.000 1.000
siglip_b16_512 768 0.9074 0.9625 0.9876 1.000 1.000
siglip_l16_256 1024 0.9100 0.9653 0.9894 0.9999 1.000
siglip_l16_384 1024 0.9072 0.9641 0.9889 0.9999 1.000
siglip_so400m_384 1152 0.8835 0.9483 0.9801 0.9994 1.0000
vit_b16_21k 768 0.7695 0.8746 0.9406 1.000 1.000
vit_l16_21k 1024 0.7142 0.8230 0.8958 0.9741 1.000
vit_s16_21k 384 0.9218 1.000 1.000 1.000 1.000
=================================================================
SCAN COMPLETE
=================================================================