anhtld commited on
Commit
c2a6e28
·
verified ·
1 Parent(s): c86c41e

refresh CTT proxy gate artifacts

Browse files
runs/ctt_val_proxy_comparison/command.txt CHANGED
@@ -1 +1 @@
1
- python scripts/build_ctt_proxy_comparison.py --out-dir runs/ctt_val_proxy_comparison
 
1
+ python scripts/build_ctt_proxy_comparison.py --out-dir runs/ctt_val_proxy_comparison --no-markdown-report
runs/ctt_val_proxy_comparison/metrics.json CHANGED
@@ -6,12 +6,15 @@
6
  {
7
  "beats_local_atlas_support": false,
8
  "candidate_diversity": 0.821928094090022,
 
9
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
 
10
  "mean_positive_distance": 0.7203264851945577,
11
  "method": "local_atlas",
12
  "negative_near_0p20": 0.03683574879227053,
13
  "negative_near_0p40": 0.20513731437644483,
14
  "num_rows": 69,
 
15
  "pptc_0p20": 0.4057971014492754,
16
  "pptc_0p40": 0.6811594202898551,
17
  "proxy_gate_pass": false,
@@ -24,12 +27,15 @@
24
  {
25
  "beats_local_atlas_support": false,
26
  "candidate_diversity": 0.9751103010950768,
 
27
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
 
28
  "mean_positive_distance": 0.9615657234367121,
29
  "method": "task_memory",
30
  "negative_near_0p20": 0.017451690821256038,
31
  "negative_near_0p40": 0.1108695652173913,
32
  "num_rows": 69,
 
33
  "pptc_0p20": 0.3188405797101449,
34
  "pptc_0p40": 0.5362318840579711,
35
  "proxy_gate_pass": false,
@@ -42,12 +48,15 @@
42
  {
43
  "beats_local_atlas_support": true,
44
  "candidate_diversity": 0.27027754304306056,
 
45
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
 
46
  "mean_positive_distance": 0.45091308311397055,
47
  "method": "ctt_residual_full",
48
  "negative_near_0p20": 0.029584063279715453,
49
  "negative_near_0p40": 0.289941117658509,
50
  "num_rows": 207,
 
51
  "pptc_0p20": 0.19806763285024154,
52
  "pptc_0p40": 0.608695652173913,
53
  "proxy_gate_pass": true,
@@ -60,12 +69,15 @@
60
  {
61
  "beats_local_atlas_support": true,
62
  "candidate_diversity": 0.24708366700914647,
 
63
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
 
64
  "mean_positive_distance": 0.44291597778605085,
65
  "method": "ctt_residual_base_context",
66
  "negative_near_0p20": 0.018245341614906832,
67
  "negative_near_0p40": 0.28596910336040765,
68
  "num_rows": 69,
 
69
  "pptc_0p20": 0.17391304347826086,
70
  "pptc_0p40": 0.6231884057971014,
71
  "proxy_gate_pass": true,
@@ -78,12 +90,15 @@
78
  {
79
  "beats_local_atlas_support": true,
80
  "candidate_diversity": 0.23968052668773945,
 
81
  "data_hash": "1d15143588697e89f7fc2f6375b2745e4d479c726c9a947ff7eb7b2705280e1a",
 
82
  "mean_positive_distance": 0.43468703005975967,
83
  "method": "ctt_residual_base_context_obs",
84
  "negative_near_0p20": 0.034336456619065314,
85
  "negative_near_0p40": 0.30197859708729274,
86
  "num_rows": 207,
 
87
  "pptc_0p20": 0.24637681159420288,
88
  "pptc_0p40": 0.6425120772946861,
89
  "proxy_gate_pass": true,
@@ -96,12 +111,15 @@
96
  {
97
  "beats_local_atlas_support": true,
98
  "candidate_diversity": 0.24174490402345686,
 
99
  "data_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
 
100
  "mean_positive_distance": 0.43402743826233176,
101
  "method": "ctt_residual_base_context_obj",
102
  "negative_near_0p20": 0.038024455415759766,
103
  "negative_near_0p40": 0.30944174585478934,
104
  "num_rows": 207,
 
105
  "pptc_0p20": 0.22705314009661837,
106
  "pptc_0p40": 0.6425120772946861,
107
  "proxy_gate_pass": true,
@@ -114,12 +132,15 @@
114
  {
115
  "beats_local_atlas_support": true,
116
  "candidate_diversity": 0.2447676812446944,
 
117
  "data_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
 
118
  "mean_positive_distance": 0.4429413077106721,
119
  "method": "ctt_residual_base_context_obs_obj",
120
  "negative_near_0p20": 0.02853647962343615,
121
  "negative_near_0p40": 0.2934249526640831,
122
  "num_rows": 207,
 
123
  "pptc_0p20": 0.20772946859903382,
124
  "pptc_0p40": 0.642512077294686,
125
  "proxy_gate_pass": true,
@@ -132,12 +153,15 @@
132
  {
133
  "beats_local_atlas_support": true,
134
  "candidate_diversity": 0.1163743674776266,
 
135
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
 
136
  "mean_positive_distance": 0.4336709210379143,
137
  "method": "ctt_gated_residual_full",
138
  "negative_near_0p20": 0.05269106899541682,
139
  "negative_near_0p40": 0.33921316203924895,
140
  "num_rows": 207,
 
141
  "pptc_0p20": 0.2318840579710145,
142
  "pptc_0p40": 0.6135265700483091,
143
  "proxy_gate_pass": false,
@@ -149,5 +173,6 @@
149
  }
150
  ],
151
  "safety_slack": 0.01,
 
152
  "split_hash": "9d1dc1c10868f76fa080895df0f2730c173f12b6db14932599d55c2b60859381"
153
  }
 
6
  {
7
  "beats_local_atlas_support": false,
8
  "candidate_diversity": 0.821928094090022,
9
+ "collapse_rate": 0.06605580409928237,
10
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
11
+ "mean_negative_distance": 0.6789322236382422,
12
  "mean_positive_distance": 0.7203264851945577,
13
  "method": "local_atlas",
14
  "negative_near_0p20": 0.03683574879227053,
15
  "negative_near_0p40": 0.20513731437644483,
16
  "num_rows": 69,
17
+ "pos_closer_than_neg": 0.5998446814623285,
18
  "pptc_0p20": 0.4057971014492754,
19
  "pptc_0p40": 0.6811594202898551,
20
  "proxy_gate_pass": false,
 
27
  {
28
  "beats_local_atlas_support": false,
29
  "candidate_diversity": 0.9751103010950768,
30
+ "collapse_rate": 0.06256038647342994,
31
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
32
+ "mean_negative_distance": 0.9219133529525894,
33
  "mean_positive_distance": 0.9615657234367121,
34
  "method": "task_memory",
35
  "negative_near_0p20": 0.017451690821256038,
36
  "negative_near_0p40": 0.1108695652173913,
37
  "num_rows": 69,
38
+ "pos_closer_than_neg": 0.5852941176470589,
39
  "pptc_0p20": 0.3188405797101449,
40
  "pptc_0p40": 0.5362318840579711,
41
  "proxy_gate_pass": false,
 
48
  {
49
  "beats_local_atlas_support": true,
50
  "candidate_diversity": 0.27027754304306056,
51
+ "collapse_rate": 0.06806780176345395,
52
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
53
+ "mean_negative_distance": 0.5291780433921461,
54
  "mean_positive_distance": 0.45091308311397055,
55
  "method": "ctt_residual_full",
56
  "negative_near_0p20": 0.029584063279715453,
57
  "negative_near_0p40": 0.289941117658509,
58
  "num_rows": 207,
59
+ "pos_closer_than_neg": 0.7351920682803036,
60
  "pptc_0p20": 0.19806763285024154,
61
  "pptc_0p40": 0.608695652173913,
62
  "proxy_gate_pass": true,
 
69
  {
70
  "beats_local_atlas_support": true,
71
  "candidate_diversity": 0.24708366700914647,
72
+ "collapse_rate": 0.06806780176345395,
73
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
74
+ "mean_negative_distance": 0.5274893750999132,
75
  "mean_positive_distance": 0.44291597778605085,
76
  "method": "ctt_residual_base_context",
77
  "negative_near_0p20": 0.018245341614906832,
78
  "negative_near_0p40": 0.28596910336040765,
79
  "num_rows": 69,
80
+ "pos_closer_than_neg": 0.7576806363571069,
81
  "pptc_0p20": 0.17391304347826086,
82
  "pptc_0p40": 0.6231884057971014,
83
  "proxy_gate_pass": true,
 
90
  {
91
  "beats_local_atlas_support": true,
92
  "candidate_diversity": 0.23968052668773945,
93
+ "collapse_rate": 0.06806780176345395,
94
  "data_hash": "1d15143588697e89f7fc2f6375b2745e4d479c726c9a947ff7eb7b2705280e1a",
95
+ "mean_negative_distance": 0.5130644364599437,
96
  "mean_positive_distance": 0.43468703005975967,
97
  "method": "ctt_residual_base_context_obs",
98
  "negative_near_0p20": 0.034336456619065314,
99
  "negative_near_0p40": 0.30197859708729274,
100
  "num_rows": 207,
101
+ "pos_closer_than_neg": 0.7717469893940482,
102
  "pptc_0p20": 0.24637681159420288,
103
  "pptc_0p40": 0.6425120772946861,
104
  "proxy_gate_pass": true,
 
111
  {
112
  "beats_local_atlas_support": true,
113
  "candidate_diversity": 0.24174490402345686,
114
+ "collapse_rate": 0.06830470254383299,
115
  "data_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
116
+ "mean_negative_distance": 0.5216229740528893,
117
  "mean_positive_distance": 0.43402743826233176,
118
  "method": "ctt_residual_base_context_obj",
119
  "negative_near_0p20": 0.038024455415759766,
120
  "negative_near_0p40": 0.30944174585478934,
121
  "num_rows": 207,
122
+ "pos_closer_than_neg": 0.7307419976537622,
123
  "pptc_0p20": 0.22705314009661837,
124
  "pptc_0p40": 0.6425120772946861,
125
  "proxy_gate_pass": true,
 
132
  {
133
  "beats_local_atlas_support": true,
134
  "candidate_diversity": 0.2447676812446944,
135
+ "collapse_rate": 0.06830470254383299,
136
  "data_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
137
+ "mean_negative_distance": 0.5333629090899681,
138
  "mean_positive_distance": 0.4429413077106721,
139
  "method": "ctt_residual_base_context_obs_obj",
140
  "negative_near_0p20": 0.02853647962343615,
141
  "negative_near_0p40": 0.2934249526640831,
142
  "num_rows": 207,
143
+ "pos_closer_than_neg": 0.7277074505015682,
144
  "pptc_0p20": 0.20772946859903382,
145
  "pptc_0p40": 0.642512077294686,
146
  "proxy_gate_pass": true,
 
153
  {
154
  "beats_local_atlas_support": true,
155
  "candidate_diversity": 0.1163743674776266,
156
+ "collapse_rate": 0.06806780176345395,
157
  "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
158
+ "mean_negative_distance": 0.5050016174085634,
159
  "mean_positive_distance": 0.4336709210379143,
160
  "method": "ctt_gated_residual_full",
161
  "negative_near_0p20": 0.05269106899541682,
162
  "negative_near_0p40": 0.33921316203924895,
163
  "num_rows": 207,
164
+ "pos_closer_than_neg": 0.7247617850559026,
165
  "pptc_0p20": 0.2318840579710145,
166
  "pptc_0p40": 0.6135265700483091,
167
  "proxy_gate_pass": false,
 
173
  }
174
  ],
175
  "safety_slack": 0.01,
176
+ "schema_version": 2,
177
  "split_hash": "9d1dc1c10868f76fa080895df0f2730c173f12b6db14932599d55c2b60859381"
178
  }
runs/ctt_val_proxy_comparison/metrics_by_seed.json CHANGED
The diff for this file is too large to render. See raw diff
 
runs/ctt_val_proxy_comparison/metrics_by_task.json CHANGED
@@ -1 +1,498 @@
1
- {}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ctt_gated_residual_full": {
3
+ "LiftPegUpright-v1": {
4
+ "candidate_diversity": 0.05970962647106902,
5
+ "collapse_rate": 0.07311091686091686,
6
+ "mean_negative_distance": 0.3839348728146849,
7
+ "mean_positive_distance": 0.25572830294882354,
8
+ "negative_near_0p20": 0.0,
9
+ "negative_near_0p40": 0.53125,
10
+ "pos_closer_than_neg": 0.9444444444444444,
11
+ "pptc_0p20": 0.5,
12
+ "pptc_0p40": 0.8888888888888888,
13
+ "proxy_support_distance": 0.23664649313175856
14
+ },
15
+ "PickCube-v1": {
16
+ "candidate_diversity": 0.2021307744588189,
17
+ "collapse_rate": 0.10000000000000002,
18
+ "mean_negative_distance": 0.5829484689774836,
19
+ "mean_positive_distance": 0.576577907999753,
20
+ "negative_near_0p20": 0.0,
21
+ "negative_near_0p40": 0.3333333333333333,
22
+ "pos_closer_than_neg": 0.7666666666666666,
23
+ "pptc_0p20": 0.0,
24
+ "pptc_0p40": 0.3333333333333333,
25
+ "proxy_support_distance": 0.4443502650555026
26
+ },
27
+ "PullCube-v1": {
28
+ "candidate_diversity": 0.14847138063951565,
29
+ "collapse_rate": 0.06404465632406808,
30
+ "mean_negative_distance": 0.7041678688460474,
31
+ "mean_positive_distance": 0.592154236746947,
32
+ "negative_near_0p20": 0.0,
33
+ "negative_near_0p40": 0.05861928104575163,
34
+ "pos_closer_than_neg": 0.6517195767195768,
35
+ "pptc_0p20": 0.029411764705882353,
36
+ "pptc_0p40": 0.3137254901960784,
37
+ "proxy_support_distance": 0.5373627305388711
38
+ },
39
+ "PushCube-v1": {
40
+ "candidate_diversity": 0.05417944485503618,
41
+ "collapse_rate": 0.06478937728937728,
42
+ "mean_negative_distance": 0.3979735192222798,
43
+ "mean_positive_distance": 0.2811870120807798,
44
+ "negative_near_0p20": 0.08974358974358974,
45
+ "negative_near_0p40": 0.44871794871794873,
46
+ "pos_closer_than_neg": 0.625,
47
+ "pptc_0p20": 0.3076923076923077,
48
+ "pptc_0p40": 0.9230769230769231,
49
+ "proxy_support_distance": 0.26109739578233915
50
+ },
51
+ "StackCube-v1": {
52
+ "candidate_diversity": 0.11447220652079974,
53
+ "collapse_rate": 0.07588217338217343,
54
+ "mean_negative_distance": 0.3264807588283944,
55
+ "mean_positive_distance": 0.268244707965138,
56
+ "negative_near_0p20": 0.16460113960113962,
57
+ "negative_near_0p40": 0.8038990638990638,
58
+ "pos_closer_than_neg": 0.7895604395604395,
59
+ "pptc_0p20": 0.5333333333333333,
60
+ "pptc_0p40": 0.9333333333333333,
61
+ "proxy_support_distance": 0.2277182791599637
62
+ }
63
+ },
64
+ "ctt_residual_base_context": {
65
+ "LiftPegUpright-v1": {
66
+ "candidate_diversity": 0.14362703522993195,
67
+ "collapse_rate": 0.07311091686091686,
68
+ "mean_negative_distance": 0.41497127070807355,
69
+ "mean_positive_distance": 0.2305104051790113,
70
+ "negative_near_0p20": 0.0,
71
+ "negative_near_0p40": 0.5,
72
+ "pos_closer_than_neg": 1.0,
73
+ "pptc_0p20": 0.6666666666666666,
74
+ "pptc_0p40": 1.0,
75
+ "proxy_support_distance": 0.2014216883582193
76
+ },
77
+ "PickCube-v1": {
78
+ "candidate_diversity": 0.25660266524569825,
79
+ "collapse_rate": 0.1,
80
+ "mean_negative_distance": 0.899796341697192,
81
+ "mean_positive_distance": 0.9817664760096816,
82
+ "negative_near_0p20": 0.0,
83
+ "negative_near_0p40": 0.0,
84
+ "pos_closer_than_neg": 0.2,
85
+ "pptc_0p20": 0.0,
86
+ "pptc_0p40": 0.0,
87
+ "proxy_support_distance": 0.7855270291361975
88
+ },
89
+ "PullCube-v1": {
90
+ "candidate_diversity": 0.3216694071900426,
91
+ "collapse_rate": 0.06404465632406808,
92
+ "mean_negative_distance": 0.6926077604425354,
93
+ "mean_positive_distance": 0.5696106164927525,
94
+ "negative_near_0p20": 0.0,
95
+ "negative_near_0p40": 0.02022058823529412,
96
+ "pos_closer_than_neg": 0.6978174603174603,
97
+ "pptc_0p20": 0.029411764705882353,
98
+ "pptc_0p40": 0.38235294117647056,
99
+ "proxy_support_distance": 0.4743135938351562
100
+ },
101
+ "PushCube-v1": {
102
+ "candidate_diversity": 0.18464687502804178,
103
+ "collapse_rate": 0.06478937728937728,
104
+ "mean_negative_distance": 0.4607380704920465,
105
+ "mean_positive_distance": 0.31700052432677295,
106
+ "negative_near_0p20": 0.052884615384615384,
107
+ "negative_near_0p40": 0.3701923076923077,
108
+ "pos_closer_than_neg": 0.71875,
109
+ "pptc_0p20": 0.23076923076923078,
110
+ "pptc_0p40": 0.8461538461538461,
111
+ "proxy_support_distance": 0.267084673449308
112
+ },
113
+ "StackCube-v1": {
114
+ "candidate_diversity": 0.17288259514532145,
115
+ "collapse_rate": 0.07588217338217337,
116
+ "mean_negative_distance": 0.35211110872802126,
117
+ "mean_positive_distance": 0.3139070522101423,
118
+ "negative_near_0p20": 0.03809523809523809,
119
+ "negative_near_0p40": 0.7487912087912088,
120
+ "pos_closer_than_neg": 0.8025030525030525,
121
+ "pptc_0p20": 0.26666666666666666,
122
+ "pptc_0p40": 0.8666666666666667,
123
+ "proxy_support_distance": 0.2492223974983373
124
+ }
125
+ },
126
+ "ctt_residual_base_context_obj": {
127
+ "LiftPegUpright-v1": {
128
+ "candidate_diversity": 0.1493872636722864,
129
+ "collapse_rate": 0.07311091686091686,
130
+ "mean_negative_distance": 0.4120512868826152,
131
+ "mean_positive_distance": 0.26240904635443296,
132
+ "negative_near_0p20": 0.0,
133
+ "negative_near_0p40": 0.4647435897435897,
134
+ "pos_closer_than_neg": 0.9375,
135
+ "pptc_0p20": 0.4444444444444444,
136
+ "pptc_0p40": 0.8888888888888888,
137
+ "proxy_support_distance": 0.22436331113570077
138
+ },
139
+ "PickCube-v1": {
140
+ "candidate_diversity": 0.2693081365535955,
141
+ "collapse_rate": 0.10000000000000002,
142
+ "mean_negative_distance": 0.7942766276118407,
143
+ "mean_positive_distance": 0.8679002702981619,
144
+ "negative_near_0p20": 0.0,
145
+ "negative_near_0p40": 0.0,
146
+ "pos_closer_than_neg": 0.3333333333333333,
147
+ "pptc_0p20": 0.0,
148
+ "pptc_0p40": 0.0,
149
+ "proxy_support_distance": 0.7622612941754188
150
+ },
151
+ "PullCube-v1": {
152
+ "candidate_diversity": 0.31298127304146967,
153
+ "collapse_rate": 0.06404465632406808,
154
+ "mean_negative_distance": 0.7058399367948929,
155
+ "mean_positive_distance": 0.5663405352240012,
156
+ "negative_near_0p20": 0.0,
157
+ "negative_near_0p40": 0.03982843137254902,
158
+ "pos_closer_than_neg": 0.6993386243386243,
159
+ "pptc_0p20": 0.058823529411764705,
160
+ "pptc_0p40": 0.4019607843137255,
161
+ "proxy_support_distance": 0.47487613941275497
162
+ },
163
+ "PushCube-v1": {
164
+ "candidate_diversity": 0.18267389477352383,
165
+ "collapse_rate": 0.06478937728937728,
166
+ "mean_negative_distance": 0.45435797530595545,
167
+ "mean_positive_distance": 0.32555326496747355,
168
+ "negative_near_0p20": 0.052884615384615384,
169
+ "negative_near_0p40": 0.375,
170
+ "pos_closer_than_neg": 0.712673611111111,
171
+ "pptc_0p20": 0.15384615384615385,
172
+ "pptc_0p40": 0.8717948717948718,
173
+ "proxy_support_distance": 0.2722421007397641
174
+ },
175
+ "StackCube-v1": {
176
+ "candidate_diversity": 0.1665761829043618,
177
+ "collapse_rate": 0.07697191697191702,
178
+ "mean_negative_distance": 0.32524565684329504,
179
+ "mean_positive_distance": 0.26785120329886264,
180
+ "negative_near_0p20": 0.12907916157916158,
181
+ "negative_near_0p40": 0.8222568172568172,
182
+ "pos_closer_than_neg": 0.7281339031339031,
183
+ "pptc_0p20": 0.6,
184
+ "pptc_0p40": 0.9333333333333333,
185
+ "proxy_support_distance": 0.20483640069851286
186
+ }
187
+ },
188
+ "ctt_residual_base_context_obs": {
189
+ "LiftPegUpright-v1": {
190
+ "candidate_diversity": 0.148958579095827,
191
+ "collapse_rate": 0.07311091686091686,
192
+ "mean_negative_distance": 0.4115981169631194,
193
+ "mean_positive_distance": 0.2362653608389971,
194
+ "negative_near_0p20": 0.0,
195
+ "negative_near_0p40": 0.47061965811965817,
196
+ "pos_closer_than_neg": 0.9965277777777778,
197
+ "pptc_0p20": 0.6666666666666666,
198
+ "pptc_0p40": 1.0,
199
+ "proxy_support_distance": 0.1957123602918266
200
+ },
201
+ "PickCube-v1": {
202
+ "candidate_diversity": 0.28828954720262945,
203
+ "collapse_rate": 0.10000000000000002,
204
+ "mean_negative_distance": 0.7587970286519089,
205
+ "mean_positive_distance": 0.7901440255768358,
206
+ "negative_near_0p20": 0.0,
207
+ "negative_near_0p40": 0.0,
208
+ "pos_closer_than_neg": 0.5333333333333333,
209
+ "pptc_0p20": 0.0,
210
+ "pptc_0p40": 0.0,
211
+ "proxy_support_distance": 0.6021893630844198
212
+ },
213
+ "PullCube-v1": {
214
+ "candidate_diversity": 0.3133621097757295,
215
+ "collapse_rate": 0.06404465632406808,
216
+ "mean_negative_distance": 0.6841223846741898,
217
+ "mean_positive_distance": 0.5733947056530471,
218
+ "negative_near_0p20": 0.0,
219
+ "negative_near_0p40": 0.02839052287581699,
220
+ "pos_closer_than_neg": 0.7119047619047619,
221
+ "pptc_0p20": 0.0392156862745098,
222
+ "pptc_0p40": 0.37254901960784315,
223
+ "proxy_support_distance": 0.4821119993756999
224
+ },
225
+ "PushCube-v1": {
226
+ "candidate_diversity": 0.17226944365330818,
227
+ "collapse_rate": 0.06478937728937728,
228
+ "mean_negative_distance": 0.4515933904080564,
229
+ "mean_positive_distance": 0.31235985654167453,
230
+ "negative_near_0p20": 0.06891025641025642,
231
+ "negative_near_0p40": 0.38247863247863245,
232
+ "pos_closer_than_neg": 0.732638888888889,
233
+ "pptc_0p20": 0.3333333333333333,
234
+ "pptc_0p40": 0.9230769230769231,
235
+ "proxy_support_distance": 0.2586271514743408
236
+ },
237
+ "StackCube-v1": {
238
+ "candidate_diversity": 0.16414005465390855,
239
+ "collapse_rate": 0.07588217338217343,
240
+ "mean_negative_distance": 0.33057222184027074,
241
+ "mean_positive_distance": 0.2819713837511487,
242
+ "negative_near_0p20": 0.09822547822547821,
243
+ "negative_near_0p40": 0.8050203500203499,
244
+ "pos_closer_than_neg": 0.8023656898656899,
245
+ "pptc_0p20": 0.4888888888888889,
246
+ "pptc_0p40": 0.9111111111111111,
247
+ "proxy_support_distance": 0.22536104103963187
248
+ }
249
+ },
250
+ "ctt_residual_base_context_obs_obj": {
251
+ "LiftPegUpright-v1": {
252
+ "candidate_diversity": 0.14766830492610186,
253
+ "collapse_rate": 0.07311091686091686,
254
+ "mean_negative_distance": 0.41649342498231295,
255
+ "mean_positive_distance": 0.2643206130229132,
256
+ "negative_near_0p20": 0.0,
257
+ "negative_near_0p40": 0.4671474358974359,
258
+ "pos_closer_than_neg": 0.9861111111111112,
259
+ "pptc_0p20": 0.4444444444444444,
260
+ "pptc_0p40": 1.0,
261
+ "proxy_support_distance": 0.22818479501118874
262
+ },
263
+ "PickCube-v1": {
264
+ "candidate_diversity": 0.33757465230985995,
265
+ "collapse_rate": 0.10000000000000002,
266
+ "mean_negative_distance": 1.1075887642380198,
267
+ "mean_positive_distance": 1.1680868800298911,
268
+ "negative_near_0p20": 0.0,
269
+ "negative_near_0p40": 0.0,
270
+ "pos_closer_than_neg": 0.43333333333333335,
271
+ "pptc_0p20": 0.0,
272
+ "pptc_0p40": 0.0,
273
+ "proxy_support_distance": 0.9803242549573113
274
+ },
275
+ "PullCube-v1": {
276
+ "candidate_diversity": 0.3136336919427042,
277
+ "collapse_rate": 0.06404465632406808,
278
+ "mean_negative_distance": 0.7162658429465761,
279
+ "mean_positive_distance": 0.565417223038266,
280
+ "negative_near_0p20": 0.0,
281
+ "negative_near_0p40": 0.029411764705882353,
282
+ "pos_closer_than_neg": 0.683994708994709,
283
+ "pptc_0p20": 0.029411764705882353,
284
+ "pptc_0p40": 0.37254901960784315,
285
+ "proxy_support_distance": 0.47243182399485906
286
+ },
287
+ "PushCube-v1": {
288
+ "candidate_diversity": 0.181916451774259,
289
+ "collapse_rate": 0.06478937728937728,
290
+ "mean_negative_distance": 0.44441461966625945,
291
+ "mean_positive_distance": 0.32881952335913633,
292
+ "negative_near_0p20": 0.04487179487179487,
293
+ "negative_near_0p40": 0.3643162393162393,
294
+ "pos_closer_than_neg": 0.6822916666666666,
295
+ "pptc_0p20": 0.23076923076923078,
296
+ "pptc_0p40": 0.9230769230769231,
297
+ "proxy_support_distance": 0.27338380647453003
298
+ },
299
+ "StackCube-v1": {
300
+ "candidate_diversity": 0.17579507499334182,
301
+ "collapse_rate": 0.07697191697191702,
302
+ "mean_negative_distance": 0.3332039593498869,
303
+ "mean_positive_distance": 0.28734001912661167,
304
+ "negative_near_0p20": 0.09237891737891737,
305
+ "negative_near_0p40": 0.7804884004884005,
306
+ "pos_closer_than_neg": 0.7293905168905168,
307
+ "pptc_0p20": 0.5111111111111111,
308
+ "pptc_0p40": 0.9111111111111111,
309
+ "proxy_support_distance": 0.225556820839634
310
+ }
311
+ },
312
+ "ctt_residual_full": {
313
+ "LiftPegUpright-v1": {
314
+ "candidate_diversity": 0.16663572531428117,
315
+ "collapse_rate": 0.07311091686091686,
316
+ "mean_negative_distance": 0.41647353998567643,
317
+ "mean_positive_distance": 0.2771405265426538,
318
+ "negative_near_0p20": 0.0,
319
+ "negative_near_0p40": 0.468215811965812,
320
+ "pos_closer_than_neg": 0.90625,
321
+ "pptc_0p20": 0.4444444444444444,
322
+ "pptc_0p40": 0.8888888888888888,
323
+ "proxy_support_distance": 0.225699402965728
324
+ },
325
+ "PickCube-v1": {
326
+ "candidate_diversity": 0.2841252624025221,
327
+ "collapse_rate": 0.10000000000000002,
328
+ "mean_negative_distance": 0.7853855604157447,
329
+ "mean_positive_distance": 0.8107379646608238,
330
+ "negative_near_0p20": 0.0,
331
+ "negative_near_0p40": 0.06666666666666667,
332
+ "pos_closer_than_neg": 0.7333333333333334,
333
+ "pptc_0p20": 0.0,
334
+ "pptc_0p40": 0.3333333333333333,
335
+ "proxy_support_distance": 0.5973002537013845
336
+ },
337
+ "PullCube-v1": {
338
+ "candidate_diversity": 0.35669628571891027,
339
+ "collapse_rate": 0.06404465632406808,
340
+ "mean_negative_distance": 0.7141779455036535,
341
+ "mean_positive_distance": 0.5869452164166006,
342
+ "negative_near_0p20": 0.0,
343
+ "negative_near_0p40": 0.0196078431372549,
344
+ "pos_closer_than_neg": 0.6872354497354497,
345
+ "pptc_0p20": 0.029411764705882353,
346
+ "pptc_0p40": 0.3235294117647059,
347
+ "proxy_support_distance": 0.48532493799704673
348
+ },
349
+ "PushCube-v1": {
350
+ "candidate_diversity": 0.1897940504075825,
351
+ "collapse_rate": 0.06478937728937728,
352
+ "mean_negative_distance": 0.4571901910242416,
353
+ "mean_positive_distance": 0.33455829528179704,
354
+ "negative_near_0p20": 0.03685897435897436,
355
+ "negative_near_0p40": 0.36217948717948717,
356
+ "pos_closer_than_neg": 0.7239583333333334,
357
+ "pptc_0p20": 0.1282051282051282,
358
+ "pptc_0p40": 0.8717948717948718,
359
+ "proxy_support_distance": 0.27709473619662045
360
+ },
361
+ "StackCube-v1": {
362
+ "candidate_diversity": 0.18468096572943052,
363
+ "collapse_rate": 0.07588217338217343,
364
+ "mean_negative_distance": 0.33657300192660006,
365
+ "mean_positive_distance": 0.28893509427463027,
366
+ "negative_near_0p20": 0.10414224664224664,
367
+ "negative_near_0p40": 0.7836650386650388,
368
+ "pos_closer_than_neg": 0.7400234025234025,
369
+ "pptc_0p20": 0.5555555555555556,
370
+ "pptc_0p40": 0.9333333333333333,
371
+ "proxy_support_distance": 0.22290972126483205
372
+ }
373
+ },
374
+ "local_atlas": {
375
+ "LiftPegUpright-v1": {
376
+ "candidate_diversity": 0.5014669861510727,
377
+ "collapse_rate": 0.07219516594516594,
378
+ "mean_negative_distance": 0.4854021065936647,
379
+ "mean_positive_distance": 0.43646670032063084,
380
+ "negative_near_0p20": 0.027777777777777776,
381
+ "negative_near_0p40": 0.24219877344877347,
382
+ "pos_closer_than_neg": 0.7192460317460317,
383
+ "pptc_0p20": 0.5,
384
+ "pptc_0p40": 1.0,
385
+ "proxy_support_distance": 0.18189470652155995
386
+ },
387
+ "PickCube-v1": {
388
+ "candidate_diversity": 1.0944000325107157,
389
+ "collapse_rate": 0.1,
390
+ "mean_negative_distance": 0.6373729677105011,
391
+ "mean_positive_distance": 0.7798613153680638,
392
+ "negative_near_0p20": 0.3,
393
+ "negative_near_0p40": 0.4,
394
+ "pos_closer_than_neg": 0.6,
395
+ "pptc_0p20": 1.0,
396
+ "pptc_0p40": 1.0,
397
+ "proxy_support_distance": 0.07461591163524217
398
+ },
399
+ "PullCube-v1": {
400
+ "candidate_diversity": 1.0209711156114212,
401
+ "collapse_rate": 0.06316930618401206,
402
+ "mean_negative_distance": 0.9870468132552196,
403
+ "mean_positive_distance": 0.9460291977432415,
404
+ "negative_near_0p20": 0.003676470588235294,
405
+ "negative_near_0p40": 0.039460784313725496,
406
+ "pos_closer_than_neg": 0.6720238095238096,
407
+ "pptc_0p20": 0.058823529411764705,
408
+ "pptc_0p40": 0.38235294117647056,
409
+ "proxy_support_distance": 0.5490009012121931
410
+ },
411
+ "PushCube-v1": {
412
+ "candidate_diversity": 0.692147376654553,
413
+ "collapse_rate": 0.0641025641025641,
414
+ "mean_negative_distance": 0.5298147189077234,
415
+ "mean_positive_distance": 0.5107316898096026,
416
+ "negative_near_0p20": 0.028846153846153848,
417
+ "negative_near_0p40": 0.23237179487179485,
418
+ "pos_closer_than_neg": 0.5859375,
419
+ "pptc_0p20": 0.7692307692307693,
420
+ "pptc_0p40": 0.9230769230769231,
421
+ "proxy_support_distance": 0.21488945685319283
422
+ },
423
+ "StackCube-v1": {
424
+ "candidate_diversity": 0.593260181033124,
425
+ "collapse_rate": 0.06957264957264958,
426
+ "mean_negative_distance": 0.40728379791043173,
427
+ "mean_positive_distance": 0.4999574180225059,
428
+ "negative_near_0p20": 0.105,
429
+ "negative_near_0p40": 0.5292521367521368,
430
+ "pos_closer_than_neg": 0.4584401709401709,
431
+ "pptc_0p20": 0.8,
432
+ "pptc_0p40": 1.0,
433
+ "proxy_support_distance": 0.15391202596431447
434
+ }
435
+ },
436
+ "task_memory": {
437
+ "LiftPegUpright-v1": {
438
+ "candidate_diversity": 0.6210576200589012,
439
+ "collapse_rate": 0.0625,
440
+ "mean_negative_distance": 0.583368643118901,
441
+ "mean_positive_distance": 0.5613930072546827,
442
+ "negative_near_0p20": 0.010416666666666666,
443
+ "negative_near_0p40": 0.13541666666666666,
444
+ "pos_closer_than_neg": 0.6875,
445
+ "pptc_0p20": 0.5,
446
+ "pptc_0p40": 1.0,
447
+ "proxy_support_distance": 0.20479276324345116
448
+ },
449
+ "PickCube-v1": {
450
+ "candidate_diversity": 1.2387041849339937,
451
+ "collapse_rate": 0.06666666666666667,
452
+ "mean_negative_distance": 0.8095929802695191,
453
+ "mean_positive_distance": 0.9291008275018162,
454
+ "negative_near_0p20": 0.26666666666666666,
455
+ "negative_near_0p40": 0.4,
456
+ "pos_closer_than_neg": 0.6,
457
+ "pptc_0p20": 1.0,
458
+ "pptc_0p40": 1.0,
459
+ "proxy_support_distance": 0.07461591163524217
460
+ },
461
+ "PullCube-v1": {
462
+ "candidate_diversity": 1.3442506282390456,
463
+ "collapse_rate": 0.0625,
464
+ "mean_negative_distance": 1.383142918984482,
465
+ "mean_positive_distance": 1.3657646662805298,
466
+ "negative_near_0p20": 0.0,
467
+ "negative_near_0p40": 0.0,
468
+ "pos_closer_than_neg": 0.4732142857142857,
469
+ "pptc_0p20": 0.029411764705882353,
470
+ "pptc_0p40": 0.08823529411764706,
471
+ "proxy_support_distance": 0.8291223046675316
472
+ },
473
+ "PushCube-v1": {
474
+ "candidate_diversity": 0.6719745815075286,
475
+ "collapse_rate": 0.0625,
476
+ "mean_negative_distance": 0.7351552918288785,
477
+ "mean_positive_distance": 0.6373056713998045,
478
+ "negative_near_0p20": 0.014423076923076924,
479
+ "negative_near_0p40": 0.07211538461538461,
480
+ "pos_closer_than_neg": 0.609375,
481
+ "pptc_0p20": 0.38461538461538464,
482
+ "pptc_0p40": 0.9230769230769231,
483
+ "proxy_support_distance": 0.26447355510409054
484
+ },
485
+ "StackCube-v1": {
486
+ "candidate_diversity": 0.5251579967031662,
487
+ "collapse_rate": 0.0625,
488
+ "mean_negative_distance": 0.5187021685529322,
489
+ "mean_positive_distance": 0.48864024429118186,
490
+ "negative_near_0p20": 0.04583333333333333,
491
+ "negative_near_0p40": 0.36666666666666664,
492
+ "pos_closer_than_neg": 0.6875,
493
+ "pptc_0p20": 0.8,
494
+ "pptc_0p40": 1.0,
495
+ "proxy_support_distance": 0.17466840356416174
496
+ }
497
+ }
498
+ }
runs/ctt_val_proxy_comparison/table.tex CHANGED
@@ -1,15 +1,15 @@
1
  % Auto-generated by scripts/build_ctt_proxy_comparison.py
2
- \begin{tabular}{lrrrrrrrc}
3
  \toprule
4
- Method & Seeds & Rows & PPTC@0.20 & PPTC@0.40 & Neg@0.20 & MeanPosDist & Diversity & Gate \\
5
  \midrule
6
- local\_atlas & 1 & 69 & 0.4058 & 0.6812 & 0.0368 & 0.7203 & 0.8219 & baseline \\
7
- task\_memory & 1 & 69 & 0.3188 & 0.5362 & 0.0175 & 0.9616 & 0.9751 & baseline \\
8
- ctt\_residual\_full & 3 & 207 & 0.1981 & 0.6087 & 0.0296 & 0.4509 & 0.2703 & pass \\
9
- ctt\_residual\_base\_context & 1 & 69 & 0.1739 & 0.6232 & 0.0182 & 0.4429 & 0.2471 & pass \\
10
- ctt\_residual\_base\_context\_obs & 3 & 207 & 0.2464 & 0.6425 & 0.0343 & 0.4347 & 0.2397 & pass \\
11
- ctt\_residual\_base\_context\_obj & 3 & 207 & 0.2271 & 0.6425 & 0.0380 & 0.4340 & 0.2417 & pass \\
12
- ctt\_residual\_base\_context\_obs\_obj & 3 & 207 & 0.2077 & 0.6425 & 0.0285 & 0.4429 & 0.2448 & pass \\
13
- ctt\_gated\_residual\_full & 3 & 207 & 0.2319 & 0.6135 & 0.0527 & 0.4337 & 0.1164 & fail \\
14
  \bottomrule
15
  \end{tabular}
 
1
  % Auto-generated by scripts/build_ctt_proxy_comparison.py
2
+ \begin{tabular}{lrrrrrrrrrrc}
3
  \toprule
4
+ Method & Seeds & Rows & PPTC@0.20 & PPTC@0.40 & Neg@0.20 & Neg@0.40 & Pos<Neg & MeanPos & MeanNeg & Collapse & Gate \\
5
  \midrule
6
+ local\_atlas & 1 & 69 & 0.4058 & 0.6812 & 0.0368 & 0.2051 & 0.5998 & 0.7203 & 0.6789 & 0.0661 & baseline \\
7
+ task\_memory & 1 & 69 & 0.3188 & 0.5362 & 0.0175 & 0.1109 & 0.5853 & 0.9616 & 0.9219 & 0.0626 & baseline \\
8
+ ctt\_residual\_full & 3 & 207 & 0.1981 & 0.6087 & 0.0296 & 0.2899 & 0.7352 & 0.4509 & 0.5292 & 0.0681 & pass \\
9
+ ctt\_residual\_base\_context & 1 & 69 & 0.1739 & 0.6232 & 0.0182 & 0.2860 & 0.7577 & 0.4429 & 0.5275 & 0.0681 & pass \\
10
+ ctt\_residual\_base\_context\_obs & 3 & 207 & 0.2464 & 0.6425 & 0.0343 & 0.3020 & 0.7717 & 0.4347 & 0.5131 & 0.0681 & pass \\
11
+ ctt\_residual\_base\_context\_obj & 3 & 207 & 0.2271 & 0.6425 & 0.0380 & 0.3094 & 0.7307 & 0.4340 & 0.5216 & 0.0683 & pass \\
12
+ ctt\_residual\_base\_context\_obs\_obj & 3 & 207 & 0.2077 & 0.6425 & 0.0285 & 0.2934 & 0.7277 & 0.4429 & 0.5334 & 0.0683 & pass \\
13
+ ctt\_gated\_residual\_full & 3 & 207 & 0.2319 & 0.6135 & 0.0527 & 0.3392 & 0.7248 & 0.4337 & 0.5050 & 0.0681 & fail \\
14
  \bottomrule
15
  \end{tabular}