Skip to content

Commit 03ae0c7

Browse files
author
Lysandre Costes
committed
SymMPO website fixes
1 parent 6a5cdc1 commit 03ae0c7

17 files changed

Lines changed: 66465 additions & 120770 deletions

‎website/assets/visualizations/muller_lyer_strong_llava_symmpo_gradcam.png‎

121 KB
Loading

‎website/css/style.css‎

Lines changed: 39 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -420,6 +420,23 @@ figcaption {
420420
margin: 0 0 0.75rem;
421421
}
422422

423+
.heas-symmpo-footnote {
424+
font-size: 0.85rem;
425+
color: var(--muted);
426+
margin: 0.5rem 0 0;
427+
}
428+
429+
.heas-header.symmpo-col,
430+
.heas-cell.symmpo-col {
431+
outline: 1px dashed rgba(49, 130, 206, 0.45);
432+
outline-offset: -1px;
433+
}
434+
435+
.heas-gated-note {
436+
margin-top: 0;
437+
margin-bottom: 0.75rem;
438+
}
439+
423440
.heas-mode-tabs {
424441
margin-bottom: 1rem;
425442
}
@@ -775,6 +792,28 @@ details.collapsible-block {
775792
display: none !important;
776793
}
777794

795+
.symmpo-heas-bar-figure {
796+
margin: 1rem 0 1.5rem;
797+
}
798+
799+
.symmpo-heas-bar-figure img {
800+
width: 100%;
801+
max-width: 720px;
802+
display: block;
803+
margin: 0 auto;
804+
}
805+
806+
.symmpo-heas-bar-figure figcaption {
807+
text-align: center;
808+
font-size: 0.9rem;
809+
color: var(--muted);
810+
margin-top: 0.5rem;
811+
}
812+
813+
.symmpo-gradcam-grid {
814+
margin: 0.75rem 0 0;
815+
}
816+
778817
.symmpo-figure-grid {
779818
display: grid;
780819
grid-template-columns: repeat(auto-fit, minmax(240px, 1fr));

‎website/data/diagnostics_summary.json‎

Lines changed: 41 additions & 41 deletions
Original file line numberDiff line numberDiff line change
@@ -39,9 +39,9 @@
3939
"n": 480,
4040
"accuracy": 0.0,
4141
"p_pred_correct": 0.0,
42-
"p_pred_illusory": 1.0,
43-
"p_pred_other": 0.0,
44-
"control_argmax_accuracy": 0.0
42+
"p_pred_illusory": 0.0,
43+
"p_pred_other": 1.0,
44+
"control_argmax_accuracy": null
4545
},
4646
{
4747
"category": "angle",
@@ -125,11 +125,11 @@
125125
"illusion_type": "zollner",
126126
"model": "llava_1.5",
127127
"n": 480,
128-
"accuracy": 0.0291666666666666,
129-
"p_pred_correct": 0.0291666666666666,
130-
"p_pred_illusory": 0.9708333333333332,
131-
"p_pred_other": 0.0,
132-
"control_argmax_accuracy": 0.10625
128+
"accuracy": 0.0,
129+
"p_pred_correct": 0.0,
130+
"p_pred_illusory": 0.0,
131+
"p_pred_other": 1.0,
132+
"control_argmax_accuracy": null
133133
},
134134
{
135135
"category": "angle",
@@ -213,11 +213,11 @@
213213
"illusion_type": "simultaneous_contrast",
214214
"model": "llava_1.5",
215215
"n": 480,
216-
"accuracy": 0.03125,
217-
"p_pred_correct": 0.03125,
218-
"p_pred_illusory": 0.5708333333333333,
219-
"p_pred_other": 0.3979166666666666,
220-
"control_argmax_accuracy": 0.1333333333333333
216+
"accuracy": 0.0,
217+
"p_pred_correct": 0.0,
218+
"p_pred_illusory": 0.0,
219+
"p_pred_other": 1.0,
220+
"control_argmax_accuracy": null
221221
},
222222
{
223223
"category": "color",
@@ -301,11 +301,11 @@
301301
"illusion_type": "whites_illusion",
302302
"model": "llava_1.5",
303303
"n": 480,
304-
"accuracy": 0.1645833333333333,
305-
"p_pred_correct": 0.1645833333333333,
306-
"p_pred_illusory": 0.3291666666666666,
307-
"p_pred_other": 0.50625,
308-
"control_argmax_accuracy": 0.2125
304+
"accuracy": 0.0,
305+
"p_pred_correct": 0.0,
306+
"p_pred_illusory": 0.0,
307+
"p_pred_other": 1.0,
308+
"control_argmax_accuracy": null
309309
},
310310
{
311311
"category": "color",
@@ -391,9 +391,9 @@
391391
"n": 480,
392392
"accuracy": 0.0,
393393
"p_pred_correct": 0.0,
394-
"p_pred_illusory": 0.5666666666666667,
395-
"p_pred_other": 0.4333333333333333,
396-
"control_argmax_accuracy": 0.0
394+
"p_pred_illusory": 0.0,
395+
"p_pred_other": 1.0,
396+
"control_argmax_accuracy": null
397397
},
398398
{
399399
"category": "geometric",
@@ -477,11 +477,11 @@
477477
"illusion_type": "muller_lyer",
478478
"model": "llava_1.5",
479479
"n": 600,
480-
"accuracy": 0.0216666666666666,
481-
"p_pred_correct": 0.0216666666666666,
482-
"p_pred_illusory": 0.8783333333333333,
483-
"p_pred_other": 0.1,
484-
"control_argmax_accuracy": 0.0316666666666666
480+
"accuracy": 0.0,
481+
"p_pred_correct": 0.0,
482+
"p_pred_illusory": 0.0,
483+
"p_pred_other": 1.0,
484+
"control_argmax_accuracy": null
485485
},
486486
{
487487
"category": "geometric",
@@ -567,9 +567,9 @@
567567
"n": 480,
568568
"accuracy": 0.0,
569569
"p_pred_correct": 0.0,
570-
"p_pred_illusory": 0.9583333333333334,
571-
"p_pred_other": 0.0416666666666666,
572-
"control_argmax_accuracy": 0.0
570+
"p_pred_illusory": 0.0,
571+
"p_pred_other": 1.0,
572+
"control_argmax_accuracy": null
573573
},
574574
{
575575
"category": "geometric",
@@ -642,10 +642,10 @@
642642
"illusion_type": "illusion_vqa",
643643
"model": "llava_1.5",
644644
"n": 435,
645-
"accuracy": 0.1471264367816092,
646-
"p_pred_correct": 0.1471264367816092,
647-
"p_pred_illusory": 0.8505747126436781,
648-
"p_pred_other": 0.0022988505747126,
645+
"accuracy": 0.15,
646+
"p_pred_correct": 0.15,
647+
"p_pred_illusory": 0.85,
648+
"p_pred_other": 0.0,
649649
"control_argmax_accuracy": null
650650
},
651651
{
@@ -697,11 +697,11 @@
697697
"illusion_type": "rotating_snakes",
698698
"model": "llava_1.5",
699699
"n": 480,
700-
"accuracy": 1.0,
701-
"p_pred_correct": 1.0,
700+
"accuracy": 0.0,
701+
"p_pred_correct": 0.0,
702702
"p_pred_illusory": 0.0,
703-
"p_pred_other": 0.0,
704-
"control_argmax_accuracy": 0.9625
703+
"p_pred_other": 1.0,
704+
"control_argmax_accuracy": null
705705
},
706706
{
707707
"category": "motion",
@@ -785,11 +785,11 @@
785785
"illusion_type": "scintillating_grid",
786786
"model": "llava_1.5",
787787
"n": 480,
788-
"accuracy": 1.0,
789-
"p_pred_correct": 1.0,
788+
"accuracy": 0.0,
789+
"p_pred_correct": 0.0,
790790
"p_pred_illusory": 0.0,
791-
"p_pred_other": 0.0,
792-
"control_argmax_accuracy": 1.0
791+
"p_pred_other": 1.0,
792+
"control_argmax_accuracy": null
793793
},
794794
{
795795
"category": "motion",

‎website/data/dpo_delta.json‎

Lines changed: 8 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -1,23 +1,23 @@
11
[
22
{
33
"category": "geometric",
4-
"llava_base": 0.5936842105263158,
4+
"llava_base": 0.12,
55
"llava_dpo": 0.88,
6-
"delta": 0.2863157894736842,
6+
"delta": 0.76,
77
"human_baseline": 0.88
88
},
99
{
1010
"category": "color",
11-
"llava_base": 0.3992771084337349,
11+
"llava_base": 0.0799999999999999,
1212
"llava_dpo": 0.92,
13-
"delta": 0.5207228915662652,
13+
"delta": 0.8400000000000001,
1414
"human_baseline": 0.92
1515
},
1616
{
1717
"category": "angle",
18-
"llava_base": 0.8654901960784314,
18+
"llava_base": 0.14,
1919
"llava_dpo": 0.8622675736961452,
20-
"delta": -0.003222622382286211,
20+
"delta": 0.7222675736961451,
2121
"human_baseline": 0.86
2222
},
2323
{
@@ -29,9 +29,9 @@
2929
},
3030
{
3131
"category": "impossible",
32-
"llava_base": 0.937574712643678,
32+
"llava_base": 0.94,
3333
"llava_dpo": 0.96516091954023,
34-
"delta": 0.027586206896552,
34+
"delta": 0.02516091954023003,
3535
"human_baseline": 0.913
3636
}
3737
]

0 commit comments

Comments
 (0)