HEART77 commited on
Commit
7c54f41
·
verified ·
1 Parent(s): 31d13fa

Publish completed frozen-v3 evaluation reports

Browse files
provenance/protocol.json ADDED
The diff for this file is too large to render. See raw diff
 
reports/frozen_v3/RESULTS.md ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # OPSD 冻结 v3 正式测试结果
2
+
3
+ 全部 147 个正式适配器及对应 Base 已评测。测试集固定,每个行为 400 题,Color/Count 各 200 题。
4
+ 沿用仓库 greedy、最多 12 个新 token、FP16 及模型专属预处理;测试结果不参与超参数或检查点选择。
5
+ 下表报告匹配效应、相对 Base 的收益和原图正确率变化,单位均为百分点。Context 收益按仓库定义反向取差。
6
+ 每个模型的原图答案单独生成,treatment/control 成对生成;保留全部逐题原始答案。
7
+
8
+ | 模型 | 阶段 | 行为 | 配置 | 匹配效应 | 收益及 95% CI | 原图正确率 | 原图正确率变化 |
9
+ |---|---|---|---|---:|---:|---:|---:|
10
+ | qwen | base_qwen | removal | Base | 84.25 | Base | 93.00 | +0.00 |
11
+ | qwen | base_qwen | replacement | Base | 87.00 | Base | 92.50 | +0.00 |
12
+ | qwen | base_qwen | context | Base | 4.25 | Base | 94.50 | +0.00 |
13
+ | internvl | base_internvl | removal | Base | 82.50 | Base | 86.50 | +0.00 |
14
+ | internvl | base_internvl | replacement | Base | 76.75 | Base | 85.50 | +0.00 |
15
+ | internvl | base_internvl | context | Base | 10.50 | Base | 85.75 | +0.00 |
16
+ | llava | base_llava | removal | Base | 79.25 | Base | 80.50 | +0.00 |
17
+ | llava | base_llava | replacement | Base | 68.50 | Base | 81.00 | +0.00 |
18
+ | llava | base_llava | context | Base | 12.50 | Base | 81.25 | +0.00 |
19
+ | qwen | joint_qwen | removal | seed13 | 94.50 | +10.25 [+6.72, +14.09] | 92.50 | -0.50 |
20
+ | qwen | joint_qwen | replacement | seed13 | 85.75 | -1.25 [-3.96, +1.46] | 91.75 | -0.75 |
21
+ | qwen | joint_qwen | context | seed13 | 5.00 | -0.75 [-3.32, +1.77] | 93.00 | -1.50 |
22
+ | qwen | joint_qwen | removal | seed42 | 93.75 | +9.50 [+5.81, +13.56] | 93.50 | +0.50 |
23
+ | qwen | joint_qwen | replacement | seed42 | 87.75 | +0.75 [-1.99, +3.53] | 92.75 | +0.25 |
24
+ | qwen | joint_qwen | context | seed42 | 5.75 | -1.50 [-4.14, +1.17] | 94.50 | +0.00 |
25
+ | qwen | joint_qwen | removal | seed73 | 94.50 | +10.25 [+6.55, +14.13] | 94.25 | +1.25 |
26
+ | qwen | joint_qwen | replacement | seed73 | 90.00 | +3.00 [+0.58, +5.49] | 94.25 | +1.75 |
27
+ | qwen | joint_qwen | context | seed73 | 5.00 | -0.75 [-3.50, +1.98] | 95.25 | +0.75 |
28
+ | internvl | fixed_internvl | removal | Deep8_s13 | 89.50 | +7.00 [+2.86, +11.15] | 87.25 | +0.75 |
29
+ | internvl | fixed_internvl | removal | Deep8_s42 | 90.25 | +7.75 [+3.79, +11.81] | 89.50 | +3.00 |
30
+ | internvl | fixed_internvl | removal | Deep8_s73 | 91.00 | +8.50 [+4.65, +12.49] | 89.00 | +2.50 |
31
+ | internvl | fixed_internvl | removal | Full_s13 | 91.25 | +8.75 [+4.58, +12.80] | 89.00 | +2.50 |
32
+ | internvl | fixed_internvl | removal | Full_s42 | 92.25 | +9.75 [+6.19, +13.50] | 89.75 | +3.25 |
33
+ | internvl | fixed_internvl | removal | Full_s73 | 91.75 | +9.25 [+5.30, +13.26] | 89.50 | +3.00 |
34
+ | internvl | fixed_internvl | removal | Middle8_s13 | 91.25 | +8.75 [+4.76, +12.83] | 89.25 | +2.75 |
35
+ | internvl | fixed_internvl | removal | Middle8_s42 | 93.25 | +10.75 [+6.98, +14.59] | 89.75 | +3.25 |
36
+ | internvl | fixed_internvl | removal | Middle8_s73 | 93.75 | +11.25 [+7.49, +15.30] | 87.50 | +1.00 |
37
+ | internvl | fixed_internvl | removal | Random8_8101_s13 | 89.75 | +7.25 [+3.03, +11.51] | 89.75 | +3.25 |
38
+ | internvl | fixed_internvl | removal | Random8_8102_s13 | 92.50 | +10.00 [+5.98, +14.08] | 88.75 | +2.25 |
39
+ | internvl | fixed_internvl | removal | Random8_8103_s13 | 92.50 | +10.00 [+6.25, +13.85] | 89.00 | +2.50 |
40
+ | internvl | fixed_internvl | removal | Random8_8104_s13 | 92.50 | +10.00 [+6.03, +14.04] | 89.75 | +3.25 |
41
+ | internvl | fixed_internvl | removal | Random8_8105_s13 | 92.50 | +10.00 [+6.05, +14.11] | 88.75 | +2.25 |
42
+ | internvl | fixed_internvl | removal | Random8_8106_s13 | 92.50 | +10.00 [+5.99, +14.25] | 89.50 | +3.00 |
43
+ | internvl | fixed_internvl | removal | Random8_8107_s13 | 93.50 | +11.00 [+7.00, +15.23] | 89.50 | +3.00 |
44
+ | internvl | fixed_internvl | removal | Random8_8108_s13 | 90.75 | +8.25 [+4.05, +12.66] | 88.25 | +1.75 |
45
+ | internvl | fixed_internvl | removal | Random8_8109_s13 | 90.50 | +8.00 [+3.81, +12.27] | 89.25 | +2.75 |
46
+ | internvl | fixed_internvl | removal | Random8_8110_s13 | 91.25 | +8.75 [+4.50, +13.07] | 88.50 | +2.00 |
47
+ | internvl | fixed_internvl | removal | Shallow8_s13 | 92.00 | +9.50 [+5.54, +14.04] | 89.50 | +3.00 |
48
+ | internvl | fixed_internvl | removal | Shallow8_s42 | 91.75 | +9.25 [+5.53, +13.23] | 90.25 | +3.75 |
49
+ | internvl | fixed_internvl | removal | Shallow8_s73 | 91.00 | +8.50 [+4.45, +12.97] | 89.25 | +2.75 |
50
+ | internvl | fixed_internvl | removal | Top8_s13 | 92.75 | +10.25 [+6.53, +14.27] | 89.75 | +3.25 |
51
+ | internvl | fixed_internvl | removal | Top8_s42 | 91.00 | +8.50 [+4.49, +12.74] | 90.50 | +4.00 |
52
+ | internvl | fixed_internvl | removal | Top8_s73 | 91.25 | +8.75 [+4.74, +13.00] | 89.25 | +2.75 |
53
+ | internvl | fixed_internvl | removal | TopDelta8_s13 | 92.00 | +9.50 [+5.74, +13.51] | 89.00 | +2.50 |
54
+ | internvl | fixed_internvl | removal | TopDelta8_s42 | 93.25 | +10.75 [+6.93, +14.75] | 89.75 | +3.25 |
55
+ | internvl | fixed_internvl | removal | TopDelta8_s73 | 92.75 | +10.25 [+6.29, +14.32] | 88.75 | +2.25 |
56
+ | internvl | fixed_internvl | replacement | Deep8_s13 | 79.75 | +3.00 [+0.13, +5.97] | 86.25 | +0.75 |
57
+ | internvl | fixed_internvl | replacement | Deep8_s42 | 79.50 | +2.75 [-0.27, +5.81] | 87.00 | +1.50 |
58
+ | internvl | fixed_internvl | replacement | Deep8_s73 | 80.25 | +3.50 [+0.22, +6.81] | 86.50 | +1.00 |
59
+ | internvl | fixed_internvl | replacement | Full_s13 | 79.75 | +3.00 [+0.16, +6.01] | 87.25 | +1.75 |
60
+ | internvl | fixed_internvl | replacement | Full_s42 | 79.25 | +2.50 [-0.68, +5.68] | 87.50 | +2.00 |
61
+ | internvl | fixed_internvl | replacement | Full_s73 | 81.00 | +4.25 [+0.55, +7.84] | 88.00 | +2.50 |
62
+ | internvl | fixed_internvl | replacement | Middle8_s13 | 79.50 | +2.75 [-0.49, +6.16] | 87.50 | +2.00 |
63
+ | internvl | fixed_internvl | replacement | Middle8_s42 | 79.25 | +2.50 [-0.89, +5.92] | 88.00 | +2.50 |
64
+ | internvl | fixed_internvl | replacement | Middle8_s73 | 80.75 | +4.00 [+0.53, +7.50] | 85.50 | +0.00 |
65
+ | internvl | fixed_internvl | replacement | Random8_8101_s13 | 79.25 | +2.50 [-0.50, +5.52] | 87.25 | +1.75 |
66
+ | internvl | fixed_internvl | replacement | Random8_8102_s13 | 80.50 | +3.75 [+0.67, +7.03] | 88.00 | +2.50 |
67
+ | internvl | fixed_internvl | replacement | Random8_8103_s13 | 79.50 | +2.75 [-0.27, +5.87] | 88.50 | +3.00 |
68
+ | internvl | fixed_internvl | replacement | Random8_8104_s13 | 79.25 | +2.50 [-0.48, +5.49] | 87.75 | +2.25 |
69
+ | internvl | fixed_internvl | replacement | Random8_8105_s13 | 80.00 | +3.25 [+0.44, +6.25] | 87.25 | +1.75 |
70
+ | internvl | fixed_internvl | replacement | Random8_8106_s13 | 79.25 | +2.50 [-0.56, +5.63] | 87.25 | +1.75 |
71
+ | internvl | fixed_internvl | replacement | Random8_8107_s13 | 79.00 | +2.25 [-0.71, +5.40] | 87.50 | +2.00 |
72
+ | internvl | fixed_internvl | replacement | Random8_8108_s13 | 79.75 | +3.00 [+0.04, +6.02] | 87.75 | +2.25 |
73
+ | internvl | fixed_internvl | replacement | Random8_8109_s13 | 80.25 | +3.50 [+0.46, +6.70] | 86.75 | +1.25 |
74
+ | internvl | fixed_internvl | replacement | Random8_8110_s13 | 80.00 | +3.25 [+0.26, +6.27] | 88.00 | +2.50 |
75
+ | internvl | fixed_internvl | replacement | Shallow8_s13 | 78.50 | +1.75 [-1.71, +5.02] | 88.00 | +2.50 |
76
+ | internvl | fixed_internvl | replacement | Shallow8_s42 | 80.00 | +3.25 [+0.01, +6.53] | 89.25 | +3.75 |
77
+ | internvl | fixed_internvl | replacement | Shallow8_s73 | 79.00 | +2.25 [-1.45, +5.85] | 88.75 | +3.25 |
78
+ | internvl | fixed_internvl | replacement | Top8_s13 | 79.00 | +2.25 [-1.00, +5.47] | 88.25 | +2.75 |
79
+ | internvl | fixed_internvl | replacement | Top8_s42 | 79.00 | +2.25 [-1.00, +5.52] | 88.00 | +2.50 |
80
+ | internvl | fixed_internvl | replacement | Top8_s73 | 79.25 | +2.50 [-1.38, +6.26] | 88.25 | +2.75 |
81
+ | internvl | fixed_internvl | replacement | TopDelta8_s13 | 80.00 | +3.25 [+0.25, +6.38] | 87.00 | +1.50 |
82
+ | internvl | fixed_internvl | replacement | TopDelta8_s42 | 79.25 | +2.50 [-0.67, +5.70] | 87.00 | +1.50 |
83
+ | internvl | fixed_internvl | replacement | TopDelta8_s73 | 80.50 | +3.75 [+0.21, +7.23] | 85.50 | +0.00 |
84
+ | internvl | fixed_internvl | context | Deep8_s13 | 7.50 | +3.00 [-0.71, +7.08] | 88.75 | +3.00 |
85
+ | internvl | fixed_internvl | context | Deep8_s42 | 7.00 | +3.50 [-0.21, +7.29] | 89.25 | +3.50 |
86
+ | internvl | fixed_internvl | context | Deep8_s73 | 7.00 | +3.50 [-0.22, +7.23] | 88.75 | +3.00 |
87
+ | internvl | fixed_internvl | context | Full_s13 | 7.75 | +2.75 [-1.04, +6.68] | 89.25 | +3.50 |
88
+ | internvl | fixed_internvl | context | Full_s42 | 8.50 | +2.00 [-1.73, +5.98] | 88.25 | +2.50 |
89
+ | internvl | fixed_internvl | context | Full_s73 | 8.00 | +2.50 [-0.91, +5.96] | 89.50 | +3.75 |
90
+ | internvl | fixed_internvl | context | Middle8_s13 | 8.75 | +1.75 [-2.00, +5.53] | 88.00 | +2.25 |
91
+ | internvl | fixed_internvl | context | Middle8_s42 | 7.25 | +3.25 [-0.54, +7.11] | 88.25 | +2.50 |
92
+ | internvl | fixed_internvl | context | Middle8_s73 | 8.50 | +2.00 [-1.44, +5.47] | 88.25 | +2.50 |
93
+ | internvl | fixed_internvl | context | Random8_8101_s13 | 7.75 | +2.75 [-0.95, +6.64] | 89.25 | +3.50 |
94
+ | internvl | fixed_internvl | context | Random8_8102_s13 | 9.75 | +0.75 [-2.97, +4.59] | 89.00 | +3.25 |
95
+ | internvl | fixed_internvl | context | Random8_8103_s13 | 6.75 | +3.75 [+0.30, +7.32] | 89.00 | +3.25 |
96
+ | internvl | fixed_internvl | context | Random8_8104_s13 | 8.25 | +2.25 [-1.65, +6.04] | 87.75 | +2.00 |
97
+ | internvl | fixed_internvl | context | Random8_8105_s13 | 8.75 | +1.75 [-1.78, +5.55] | 89.50 | +3.75 |
98
+ | internvl | fixed_internvl | context | Random8_8106_s13 | 8.75 | +1.75 [-1.81, +5.37] | 89.50 | +3.75 |
99
+ | internvl | fixed_internvl | context | Random8_8107_s13 | 8.25 | +2.25 [-1.50, +6.01] | 89.00 | +3.25 |
100
+ | internvl | fixed_internvl | context | Random8_8108_s13 | 8.75 | +1.75 [-2.07, +5.78] | 89.00 | +3.25 |
101
+ | internvl | fixed_internvl | context | Random8_8109_s13 | 9.25 | +1.25 [-2.40, +5.09] | 89.25 | +3.50 |
102
+ | internvl | fixed_internvl | context | Random8_8110_s13 | 8.00 | +2.50 [-1.12, +6.25] | 89.25 | +3.50 |
103
+ | internvl | fixed_internvl | context | Shallow8_s13 | 5.25 | +5.25 [+1.46, +9.18] | 89.75 | +4.00 |
104
+ | internvl | fixed_internvl | context | Shallow8_s42 | 8.50 | +2.00 [-1.73, +5.86] | 89.50 | +3.75 |
105
+ | internvl | fixed_internvl | context | Shallow8_s73 | 8.25 | +2.25 [-1.68, +6.50] | 88.00 | +2.25 |
106
+ | internvl | fixed_internvl | context | Top8_s13 | 8.25 | +2.25 [-1.50, +6.02] | 87.75 | +2.00 |
107
+ | internvl | fixed_internvl | context | Top8_s42 | 7.00 | +3.50 [-0.05, +7.25] | 88.25 | +2.50 |
108
+ | internvl | fixed_internvl | context | Top8_s73 | 7.00 | +3.50 [+0.15, +7.18] | 88.25 | +2.50 |
109
+ | internvl | fixed_internvl | context | TopDelta8_s13 | 8.75 | +1.75 [-1.77, +5.47] | 87.75 | +2.00 |
110
+ | internvl | fixed_internvl | context | TopDelta8_s42 | 7.25 | +3.25 [-0.51, +7.21] | 89.00 | +3.25 |
111
+ | internvl | fixed_internvl | context | TopDelta8_s73 | 7.75 | +2.75 [-1.06, +6.75] | 89.25 | +3.50 |
112
+ | internvl | single_internvl | removal | L01_s13 | 84.75 | +2.25 [-0.81, +5.42] | 89.75 | +3.25 |
113
+ | internvl | single_internvl | removal | L02_s13 | 84.75 | +2.25 [-0.74, +5.35] | 90.00 | +3.50 |
114
+ | internvl | single_internvl | removal | L07_s13 | 89.25 | +6.75 [+2.76, +10.76] | 88.75 | +2.25 |
115
+ | internvl | single_internvl | removal | L09_s13 | 88.00 | +5.50 [+1.67, +9.48] | 89.50 | +3.00 |
116
+ | internvl | single_internvl | removal | L11_s13 | 89.75 | +7.25 [+3.73, +10.94] | 89.25 | +2.75 |
117
+ | internvl | single_internvl | removal | L16_s13 | 89.75 | +7.25 [+3.47, +11.42] | 89.50 | +3.00 |
118
+ | internvl | single_internvl | removal | L20_s13 | 91.00 | +8.50 [+4.99, +12.47] | 89.50 | +3.00 |
119
+ | internvl | single_internvl | removal | L25_s13 | 89.75 | +7.25 [+3.16, +11.55] | 88.75 | +2.25 |
120
+ | internvl | single_internvl | removal | L29_s13 | 90.00 | +7.50 [+3.54, +11.54] | 89.50 | +3.00 |
121
+ | internvl | single_internvl | removal | L34_s13 | 88.75 | +6.25 [+2.64, +9.89] | 88.75 | +2.25 |
122
+ | internvl | single_internvl | replacement | L01_s13 | 77.50 | +0.75 [-1.90, +3.33] | 87.25 | +1.75 |
123
+ | internvl | single_internvl | replacement | L02_s13 | 77.50 | +0.75 [-1.76, +3.34] | 87.25 | +1.75 |
124
+ | internvl | single_internvl | replacement | L07_s13 | 78.25 | +1.50 [-1.27, +4.27] | 88.25 | +2.75 |
125
+ | internvl | single_internvl | replacement | L09_s13 | 78.50 | +1.75 [-1.12, +4.71] | 88.00 | +2.50 |
126
+ | internvl | single_internvl | replacement | L11_s13 | 78.00 | +1.25 [-1.68, +4.21] | 88.00 | +2.50 |
127
+ | internvl | single_internvl | replacement | L16_s13 | 80.00 | +3.25 [+0.08, +6.52] | 88.50 | +3.00 |
128
+ | internvl | single_internvl | replacement | L20_s13 | 80.00 | +3.25 [-0.00, +6.63] | 88.00 | +2.50 |
129
+ | internvl | single_internvl | replacement | L25_s13 | 79.25 | +2.50 [-0.61, +5.66] | 86.75 | +1.25 |
130
+ | internvl | single_internvl | replacement | L29_s13 | 77.75 | +1.00 [-1.97, +3.90] | 87.50 | +2.00 |
131
+ | internvl | single_internvl | replacement | L34_s13 | 78.25 | +1.50 [-1.07, +4.10] | 87.00 | +1.50 |
132
+ | internvl | single_internvl | context | L01_s13 | 7.75 | +2.75 [-0.98, +6.56] | 87.25 | +1.50 |
133
+ | internvl | single_internvl | context | L02_s13 | 7.25 | +3.25 [-0.12, +6.81] | 86.75 | +1.00 |
134
+ | internvl | single_internvl | context | L07_s13 | 8.00 | +2.50 [-1.28, +6.48] | 88.50 | +2.75 |
135
+ | internvl | single_internvl | context | L09_s13 | 8.75 | +1.75 [-1.71, +5.44] | 88.75 | +3.00 |
136
+ | internvl | single_internvl | context | L11_s13 | 9.50 | +1.00 [-2.94, +5.13] | 88.00 | +2.25 |
137
+ | internvl | single_internvl | context | L16_s13 | 8.25 | +2.25 [-1.11, +6.11] | 88.00 | +2.25 |
138
+ | internvl | single_internvl | context | L20_s13 | 9.00 | +1.50 [-1.98, +5.11] | 88.75 | +3.00 |
139
+ | internvl | single_internvl | context | L25_s13 | 8.25 | +2.25 [-1.29, +5.92] | 88.25 | +2.50 |
140
+ | internvl | single_internvl | context | L29_s13 | 9.25 | +1.25 [-2.25, +5.00] | 88.50 | +2.75 |
141
+ | internvl | single_internvl | context | L34_s13 | 7.75 | +2.75 [-1.14, +6.98] | 88.00 | +2.25 |
142
+ | llava | single_llava | removal | L01_s13 | 88.50 | +9.25 [+5.14, +13.64] | 78.75 | -1.75 |
143
+ | llava | single_llava | removal | L02_s13 | 85.50 | +6.25 [+1.61, +11.24] | 75.50 | -5.00 |
144
+ | llava | single_llava | removal | L06_s13 | 89.50 | +10.25 [+6.22, +14.49] | 80.00 | -0.50 |
145
+ | llava | single_llava | removal | L08_s13 | 88.50 | +9.25 [+5.16, +13.59] | 79.25 | -1.25 |
146
+ | llava | single_llava | removal | L10_s13 | 90.50 | +11.25 [+7.12, +15.67] | 79.50 | -1.00 |
147
+ | llava | single_llava | removal | L14_s13 | 89.75 | +10.50 [+6.45, +14.98] | 81.00 | +0.50 |
148
+ | llava | single_llava | removal | L18_s13 | 88.00 | +8.75 [+4.37, +13.04] | 82.50 | +2.00 |
149
+ | llava | single_llava | removal | L22_s13 | 89.75 | +10.50 [+6.50, +14.93] | 81.50 | +1.00 |
150
+ | llava | single_llava | removal | L26_s13 | 85.75 | +6.50 [+2.63, +10.55] | 81.75 | +1.25 |
151
+ | llava | single_llava | removal | L30_s13 | 89.00 | +9.75 [+5.61, +14.28] | 83.75 | +3.25 |
152
+ | llava | single_llava | replacement | L01_s13 | 67.50 | -1.00 [-4.98, +3.05] | 77.00 | -4.00 |
153
+ | llava | single_llava | replacement | L02_s13 | 67.00 | -1.50 [-5.38, +2.71] | 74.75 | -6.25 |
154
+ | llava | single_llava | replacement | L06_s13 | 66.50 | -2.00 [-6.18, +2.30] | 76.25 | -4.75 |
155
+ | llava | single_llava | replacement | L08_s13 | 65.00 | -3.50 [-7.51, +0.61] | 74.00 | -7.00 |
156
+ | llava | single_llava | replacement | L10_s13 | 69.00 | +0.50 [-3.57, +4.71] | 77.50 | -3.50 |
157
+ | llava | single_llava | replacement | L14_s13 | 65.00 | -3.50 [-7.80, +0.80] | 75.75 | -5.25 |
158
+ | llava | single_llava | replacement | L18_s13 | 67.00 | -1.50 [-5.49, +2.68] | 75.75 | -5.25 |
159
+ | llava | single_llava | replacement | L22_s13 | 66.00 | -2.50 [-6.61, +1.80] | 73.50 | -7.50 |
160
+ | llava | single_llava | replacement | L26_s13 | 67.75 | -0.75 [-4.55, +3.29] | 76.75 | -4.25 |
161
+ | llava | single_llava | replacement | L30_s13 | 67.75 | -0.75 [-3.98, +2.64] | 78.00 | -3.00 |
162
+ | llava | single_llava | context | L01_s13 | 10.50 | +2.00 [-2.05, +6.10] | 81.75 | +0.50 |
163
+ | llava | single_llava | context | L02_s13 | 9.50 | +3.00 [-1.02, +7.17] | 81.75 | +0.50 |
164
+ | llava | single_llava | context | L06_s13 | 9.50 | +3.00 [-1.25, +7.15] | 81.25 | +0.00 |
165
+ | llava | single_llava | context | L08_s13 | 8.75 | +3.75 [-0.49, +8.07] | 82.25 | +1.00 |
166
+ | llava | single_llava | context | L10_s13 | 12.25 | +0.25 [-4.25, +4.60] | 81.00 | -0.25 |
167
+ | llava | single_llava | context | L14_s13 | 9.50 | +3.00 [-1.25, +7.21] | 82.25 | +1.00 |
168
+ | llava | single_llava | context | L18_s13 | 9.50 | +3.00 [-1.14, +7.46] | 83.00 | +1.75 |
169
+ | llava | single_llava | context | L22_s13 | 8.75 | +3.75 [-0.48, +8.19] | 81.75 | +0.50 |
170
+ | llava | single_llava | context | L26_s13 | 9.75 | +2.75 [-1.20, +6.96] | 82.75 | +1.50 |
171
+ | llava | single_llava | context | L30_s13 | 13.25 | -0.75 [-4.25, +2.75] | 84.25 | +3.00 |
172
+
173
+ 跨种子/随机层集合的 bootstrap、预设配置差异,以及 InternVL 定位分数与测试收益的相关性见 summary.json 和各分析 JSON。
174
+ LLaVA 没有本次训练集上的定位分数,因此不报告其定位—收益相关性。上传独立于测试完成状态。
reports/frozen_v3/fixed_internvl_context.json ADDED
@@ -0,0 +1,451 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "context",
3
+ "arms": {
4
+ "Deep8": {
5
+ "color": {
6
+ "gain": 0.6666666666666666,
7
+ "ci95": [
8
+ -2.8851495726495724,
9
+ 4.30134274862248
10
+ ]
11
+ },
12
+ "count": {
13
+ "gain": 6.0,
14
+ "ci95": [
15
+ -0.3089387464387463,
16
+ 12.440634414318627
17
+ ]
18
+ },
19
+ "behavior": {
20
+ "gain": 3.3333333333333335,
21
+ "ci95": [
22
+ -0.14629087479152567,
23
+ 6.896319368090022
24
+ ]
25
+ },
26
+ "runs": 3
27
+ },
28
+ "Full": {
29
+ "color": {
30
+ "gain": -1.3333333333333333,
31
+ "ci95": [
32
+ -5.235845713627626,
33
+ 2.525252525252525
34
+ ]
35
+ },
36
+ "count": {
37
+ "gain": 6.166666666666667,
38
+ "ci95": [
39
+ 0.1593250223139166,
40
+ 12.439365671641797
41
+ ]
42
+ },
43
+ "behavior": {
44
+ "gain": 2.4166666666666665,
45
+ "ci95": [
46
+ -1.0492949917524397,
47
+ 6.024416320892247
48
+ ]
49
+ },
50
+ "runs": 3
51
+ },
52
+ "Middle8": {
53
+ "color": {
54
+ "gain": -1.1666666666666667,
55
+ "ci95": [
56
+ -5.090369002006933,
57
+ 2.551677887649907
58
+ ]
59
+ },
60
+ "count": {
61
+ "gain": 5.833333333333333,
62
+ "ci95": [
63
+ -0.15515003152702056,
64
+ 11.800633309566754
65
+ ]
66
+ },
67
+ "behavior": {
68
+ "gain": 2.3333333333333335,
69
+ "ci95": [
70
+ -1.1875944213336809,
71
+ 5.952674244050197
72
+ ]
73
+ },
74
+ "runs": 3
75
+ },
76
+ "Random8": {
77
+ "color": {
78
+ "gain": -1.35,
79
+ "ci95": [
80
+ -5.240502450980392,
81
+ 2.376411122420653
82
+ ]
83
+ },
84
+ "count": {
85
+ "gain": 5.5,
86
+ "ci95": [
87
+ -0.5127975549332181,
88
+ 11.65011934386268
89
+ ]
90
+ },
91
+ "behavior": {
92
+ "gain": 2.075,
93
+ "ci95": [
94
+ -1.4575639045988473,
95
+ 5.653853078777224
96
+ ]
97
+ },
98
+ "runs": 10
99
+ },
100
+ "Shallow8": {
101
+ "color": {
102
+ "gain": -1.6666666666666667,
103
+ "ci95": [
104
+ -5.650520030816641,
105
+ 2.041077969649399
106
+ ]
107
+ },
108
+ "count": {
109
+ "gain": 8.0,
110
+ "ci95": [
111
+ 1.1709938581797883,
112
+ 15.13607496138552
113
+ ]
114
+ },
115
+ "behavior": {
116
+ "gain": 3.1666666666666665,
117
+ "ci95": [
118
+ -0.7516368938263859,
119
+ 7.30842545834873
120
+ ]
121
+ },
122
+ "runs": 3
123
+ },
124
+ "Top8": {
125
+ "color": {
126
+ "gain": -1.1666666666666667,
127
+ "ci95": [
128
+ -5.0,
129
+ 2.6604251737230493
130
+ ]
131
+ },
132
+ "count": {
133
+ "gain": 7.333333333333333,
134
+ "ci95": [
135
+ 1.2930001198609617,
136
+ 13.70608141733851
137
+ ]
138
+ },
139
+ "behavior": {
140
+ "gain": 3.0833333333333335,
141
+ "ci95": [
142
+ -0.38387615699588057,
143
+ 6.626760928184765
144
+ ]
145
+ },
146
+ "runs": 3
147
+ },
148
+ "TopDelta8": {
149
+ "color": {
150
+ "gain": -1.0,
151
+ "ci95": [
152
+ -5.009047829822689,
153
+ 2.7178325901152003
154
+ ]
155
+ },
156
+ "count": {
157
+ "gain": 6.166666666666667,
158
+ "ci95": [
159
+ 0.0,
160
+ 12.661406765676572
161
+ ]
162
+ },
163
+ "behavior": {
164
+ "gain": 2.5833333333333335,
165
+ "ci95": [
166
+ -1.0620537828110266,
167
+ 6.287815742174594
168
+ ]
169
+ },
170
+ "runs": 3
171
+ }
172
+ },
173
+ "contrasts": {
174
+ "Top8-Random8": {
175
+ "color": {
176
+ "difference": 0.18333333333333335,
177
+ "ci95": [
178
+ -2.160848960248397,
179
+ 2.364025682736864
180
+ ]
181
+ },
182
+ "count": {
183
+ "difference": 1.833333333333333,
184
+ "ci95": [
185
+ -1.2166732804232814,
186
+ 5.50582260267291
187
+ ]
188
+ },
189
+ "behavior": {
190
+ "difference": 1.0083333333333333,
191
+ "ci95": [
192
+ -0.8398903832366266,
193
+ 2.9564089965896843
194
+ ]
195
+ }
196
+ },
197
+ "TopDelta8-Random8": {
198
+ "color": {
199
+ "difference": 0.3500000000000001,
200
+ "ci95": [
201
+ -1.4716913468049464,
202
+ 2.1976762207848584
203
+ ]
204
+ },
205
+ "count": {
206
+ "difference": 0.666666666666667,
207
+ "ci95": [
208
+ -2.492174857724686,
209
+ 4.010644745814078
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "difference": 0.5083333333333333,
214
+ "ci95": [
215
+ -1.4081065709933367,
216
+ 2.507500587254224
217
+ ]
218
+ }
219
+ },
220
+ "Shallow8-Random8": {
221
+ "color": {
222
+ "difference": -0.31666666666666665,
223
+ "ci95": [
224
+ -3.317155993278566,
225
+ 2.921225885475147
226
+ ]
227
+ },
228
+ "count": {
229
+ "difference": 2.5,
230
+ "ci95": [
231
+ -2.0037219970730615,
232
+ 7.081490686115112
233
+ ]
234
+ },
235
+ "behavior": {
236
+ "difference": 1.0916666666666663,
237
+ "ci95": [
238
+ -1.890742755680293,
239
+ 4.073317216367707
240
+ ]
241
+ }
242
+ },
243
+ "Middle8-Random8": {
244
+ "color": {
245
+ "difference": 0.18333333333333335,
246
+ "ci95": [
247
+ -2.1770606300772735,
248
+ 2.1987757514353268
249
+ ]
250
+ },
251
+ "count": {
252
+ "difference": 0.33333333333333304,
253
+ "ci95": [
254
+ -2.908586511527689,
255
+ 3.6753663003663
256
+ ]
257
+ },
258
+ "behavior": {
259
+ "difference": 0.2583333333333333,
260
+ "ci95": [
261
+ -1.876531954887218,
262
+ 2.247400161900185
263
+ ]
264
+ }
265
+ },
266
+ "Deep8-Random8": {
267
+ "color": {
268
+ "difference": 2.0166666666666666,
269
+ "ci95": [
270
+ -0.862992179997256,
271
+ 5.200811991162236
272
+ ]
273
+ },
274
+ "count": {
275
+ "difference": 0.5,
276
+ "ci95": [
277
+ -3.760459056186194,
278
+ 4.877094393164028
279
+ ]
280
+ },
281
+ "behavior": {
282
+ "difference": 1.2583333333333333,
283
+ "ci95": [
284
+ -1.2798553307024108,
285
+ 3.905312345388439
286
+ ]
287
+ }
288
+ },
289
+ "Top8-Shallow8": {
290
+ "color": {
291
+ "difference": 0.5,
292
+ "ci95": [
293
+ -2.623659844054581,
294
+ 3.4017966160823323
295
+ ]
296
+ },
297
+ "count": {
298
+ "difference": -0.666666666666667,
299
+ "ci95": [
300
+ -5.664223219658161,
301
+ 4.592425431711148
302
+ ]
303
+ },
304
+ "behavior": {
305
+ "difference": -0.08333333333333304,
306
+ "ci95": [
307
+ -3.2126736137582186,
308
+ 2.9376729918440287
309
+ ]
310
+ }
311
+ },
312
+ "Deep8-Full": {
313
+ "color": {
314
+ "difference": 2.0,
315
+ "ci95": [
316
+ -0.9760951340856691,
317
+ 5.263381668456858
318
+ ]
319
+ },
320
+ "count": {
321
+ "difference": -0.16666666666666696,
322
+ "ci95": [
323
+ -4.598006847640008,
324
+ 4.146455223880598
325
+ ]
326
+ },
327
+ "behavior": {
328
+ "difference": 0.916666666666667,
329
+ "ci95": [
330
+ -1.5576810891528263,
331
+ 3.418970352564103
332
+ ]
333
+ }
334
+ },
335
+ "Middle8-Full": {
336
+ "color": {
337
+ "difference": 0.16666666666666652,
338
+ "ci95": [
339
+ -2.427480464125029,
340
+ 2.5516571362844678
341
+ ]
342
+ },
343
+ "count": {
344
+ "difference": -0.3333333333333339,
345
+ "ci95": [
346
+ -4.085338680926918,
347
+ 3.2815836715247113
348
+ ]
349
+ },
350
+ "behavior": {
351
+ "difference": -0.08333333333333304,
352
+ "ci95": [
353
+ -2.33340673272612,
354
+ 1.9279814647004245
355
+ ]
356
+ }
357
+ },
358
+ "Random8-Full": {
359
+ "color": {
360
+ "difference": -0.01666666666666683,
361
+ "ci95": [
362
+ -1.8790123917587334,
363
+ 1.7173896323642523
364
+ ]
365
+ },
366
+ "count": {
367
+ "difference": -0.666666666666667,
368
+ "ci95": [
369
+ -3.688064942288972,
370
+ 2.397506105540414
371
+ ]
372
+ },
373
+ "behavior": {
374
+ "difference": -0.34166666666666634,
375
+ "ci95": [
376
+ -1.9993687147141845,
377
+ 1.3131350530382426
378
+ ]
379
+ }
380
+ },
381
+ "Shallow8-Full": {
382
+ "color": {
383
+ "difference": -0.3333333333333335,
384
+ "ci95": [
385
+ -3.465777547904044,
386
+ 3.0748924031628486
387
+ ]
388
+ },
389
+ "count": {
390
+ "difference": 1.833333333333333,
391
+ "ci95": [
392
+ -3.151305970149254,
393
+ 7.193224896411884
394
+ ]
395
+ },
396
+ "behavior": {
397
+ "difference": 0.75,
398
+ "ci95": [
399
+ -2.329369379284607,
400
+ 3.9194275271084056
401
+ ]
402
+ }
403
+ },
404
+ "Top8-Full": {
405
+ "color": {
406
+ "difference": 0.16666666666666652,
407
+ "ci95": [
408
+ -2.39298493525382,
409
+ 2.4756716325838206
410
+ ]
411
+ },
412
+ "count": {
413
+ "difference": 1.166666666666666,
414
+ "ci95": [
415
+ -2.469596462133776,
416
+ 5.1158288905813665
417
+ ]
418
+ },
419
+ "behavior": {
420
+ "difference": 0.666666666666667,
421
+ "ci95": [
422
+ -1.4385378065395786,
423
+ 2.642019828518
424
+ ]
425
+ }
426
+ },
427
+ "TopDelta8-Full": {
428
+ "color": {
429
+ "difference": 0.33333333333333326,
430
+ "ci95": [
431
+ -1.5708872967759713,
432
+ 2.1835347654124955
433
+ ]
434
+ },
435
+ "count": {
436
+ "difference": 0.0,
437
+ "ci95": [
438
+ -3.61292656359232,
439
+ 3.6088003130209954
440
+ ]
441
+ },
442
+ "behavior": {
443
+ "difference": 0.16666666666666696,
444
+ "ci95": [
445
+ -1.8426120250215976,
446
+ 2.082723897389739
447
+ ]
448
+ }
449
+ }
450
+ }
451
+ }
reports/frozen_v3/fixed_internvl_removal.json ADDED
@@ -0,0 +1,451 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "removal",
3
+ "arms": {
4
+ "Deep8": {
5
+ "color": {
6
+ "gain": 10.166666666666666,
7
+ "ci95": [
8
+ 4.537149096266674,
9
+ 15.781662433388119
10
+ ]
11
+ },
12
+ "count": {
13
+ "gain": 5.333333333333333,
14
+ "ci95": [
15
+ 0.0,
16
+ 10.72728106143948
17
+ ]
18
+ },
19
+ "behavior": {
20
+ "gain": 7.75,
21
+ "ci95": [
22
+ 3.916349296128708,
23
+ 11.591823277769754
24
+ ]
25
+ },
26
+ "runs": 3
27
+ },
28
+ "Full": {
29
+ "color": {
30
+ "gain": 9.333333333333334,
31
+ "ci95": [
32
+ 3.7233612201335817,
33
+ 15.183566882250283
34
+ ]
35
+ },
36
+ "count": {
37
+ "gain": 9.166666666666666,
38
+ "ci95": [
39
+ 4.323476942288824,
40
+ 14.361775596567533
41
+ ]
42
+ },
43
+ "behavior": {
44
+ "gain": 9.25,
45
+ "ci95": [
46
+ 5.482616495413512,
47
+ 13.075335023803161
48
+ ]
49
+ },
50
+ "runs": 3
51
+ },
52
+ "Middle8": {
53
+ "color": {
54
+ "gain": 12.166666666666666,
55
+ "ci95": [
56
+ 6.7306957186544345,
57
+ 17.894823265059202
58
+ ]
59
+ },
60
+ "count": {
61
+ "gain": 8.333333333333334,
62
+ "ci95": [
63
+ 3.2783364317503243,
64
+ 13.756986362620168
65
+ ]
66
+ },
67
+ "behavior": {
68
+ "gain": 10.25,
69
+ "ci95": [
70
+ 6.360450609418865,
71
+ 14.26668681251842
72
+ ]
73
+ },
74
+ "runs": 3
75
+ },
76
+ "Random8": {
77
+ "color": {
78
+ "gain": 9.95,
79
+ "ci95": [
80
+ 3.951123234916559,
81
+ 16.048490427098677
82
+ ]
83
+ },
84
+ "count": {
85
+ "gain": 8.7,
86
+ "ci95": [
87
+ 3.9285246821241584,
88
+ 14.077718205091049
89
+ ]
90
+ },
91
+ "behavior": {
92
+ "gain": 9.325,
93
+ "ci95": [
94
+ 5.38261732847098,
95
+ 13.281660126767152
96
+ ]
97
+ },
98
+ "runs": 10
99
+ },
100
+ "Shallow8": {
101
+ "color": {
102
+ "gain": 10.833333333333334,
103
+ "ci95": [
104
+ 5.102603605634744,
105
+ 16.838625063976018
106
+ ]
107
+ },
108
+ "count": {
109
+ "gain": 7.333333333333333,
110
+ "ci95": [
111
+ 1.9999231950844854,
112
+ 12.908993036466924
113
+ ]
114
+ },
115
+ "behavior": {
116
+ "gain": 9.083333333333334,
117
+ "ci95": [
118
+ 5.151902337103394,
119
+ 13.160007402628557
120
+ ]
121
+ },
122
+ "runs": 3
123
+ },
124
+ "Top8": {
125
+ "color": {
126
+ "gain": 10.5,
127
+ "ci95": [
128
+ 4.614942528735632,
129
+ 16.26428097167846
130
+ ]
131
+ },
132
+ "count": {
133
+ "gain": 7.833333333333333,
134
+ "ci95": [
135
+ 2.897635778070561,
136
+ 13.089161057092996
137
+ ]
138
+ },
139
+ "behavior": {
140
+ "gain": 9.166666666666666,
141
+ "ci95": [
142
+ 5.217074793047767,
143
+ 13.128772781465317
144
+ ]
145
+ },
146
+ "runs": 3
147
+ },
148
+ "TopDelta8": {
149
+ "color": {
150
+ "gain": 11.833333333333334,
151
+ "ci95": [
152
+ 6.292517006802721,
153
+ 17.51297738227661
154
+ ]
155
+ },
156
+ "count": {
157
+ "gain": 8.5,
158
+ "ci95": [
159
+ 3.7037037037037037,
160
+ 13.592741320591674
161
+ ]
162
+ },
163
+ "behavior": {
164
+ "gain": 10.166666666666666,
165
+ "ci95": [
166
+ 6.272588456853881,
167
+ 13.977588910057953
168
+ ]
169
+ },
170
+ "runs": 3
171
+ }
172
+ },
173
+ "contrasts": {
174
+ "Top8-Random8": {
175
+ "color": {
176
+ "difference": 0.5500000000000007,
177
+ "ci95": [
178
+ -2.181528835236759,
179
+ 3.200743853547795
180
+ ]
181
+ },
182
+ "count": {
183
+ "difference": -0.8666666666666663,
184
+ "ci95": [
185
+ -3.832101237864477,
186
+ 1.9146118721461194
187
+ ]
188
+ },
189
+ "behavior": {
190
+ "difference": -0.15833333333333321,
191
+ "ci95": [
192
+ -2.281880436956537,
193
+ 1.9427180643443083
194
+ ]
195
+ }
196
+ },
197
+ "TopDelta8-Random8": {
198
+ "color": {
199
+ "difference": 1.8833333333333346,
200
+ "ci95": [
201
+ -0.5586611938528271,
202
+ 4.410613529614792
203
+ ]
204
+ },
205
+ "count": {
206
+ "difference": -0.1999999999999993,
207
+ "ci95": [
208
+ -3.093110380116958,
209
+ 2.729087602783051
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "difference": 0.8416666666666668,
214
+ "ci95": [
215
+ -0.9637530638165153,
216
+ 2.8065518555178453
217
+ ]
218
+ }
219
+ },
220
+ "Shallow8-Random8": {
221
+ "color": {
222
+ "difference": 0.8833333333333346,
223
+ "ci95": [
224
+ -2.079517704517706,
225
+ 4.029173252050881
226
+ ]
227
+ },
228
+ "count": {
229
+ "difference": -1.3666666666666663,
230
+ "ci95": [
231
+ -4.640565650343903,
232
+ 1.6243967392663345
233
+ ]
234
+ },
235
+ "behavior": {
236
+ "difference": -0.24166666666666536,
237
+ "ci95": [
238
+ -2.327714592834495,
239
+ 1.9401803985788748
240
+ ]
241
+ }
242
+ },
243
+ "Middle8-Random8": {
244
+ "color": {
245
+ "difference": 2.216666666666667,
246
+ "ci95": [
247
+ 0.055489417989417845,
248
+ 4.628924648223668
249
+ ]
250
+ },
251
+ "count": {
252
+ "difference": -0.36666666666666536,
253
+ "ci95": [
254
+ -3.621885738831616,
255
+ 2.7726371951219524
256
+ ]
257
+ },
258
+ "behavior": {
259
+ "difference": 0.9250000000000007,
260
+ "ci95": [
261
+ -1.18797866186717,
262
+ 3.0659404566424278
263
+ ]
264
+ }
265
+ },
266
+ "Deep8-Random8": {
267
+ "color": {
268
+ "difference": 0.21666666666666679,
269
+ "ci95": [
270
+ -2.68340210355987,
271
+ 3.155196264111707
272
+ ]
273
+ },
274
+ "count": {
275
+ "difference": -3.3666666666666663,
276
+ "ci95": [
277
+ -7.067083333333334,
278
+ 0.10530537573881361
279
+ ]
280
+ },
281
+ "behavior": {
282
+ "difference": -1.5749999999999993,
283
+ "ci95": [
284
+ -3.9747666664836414,
285
+ 0.7099735526578647
286
+ ]
287
+ }
288
+ },
289
+ "Top8-Shallow8": {
290
+ "color": {
291
+ "difference": -0.3333333333333339,
292
+ "ci95": [
293
+ -3.166666666666666,
294
+ 2.475247524752475
295
+ ]
296
+ },
297
+ "count": {
298
+ "difference": 0.5,
299
+ "ci95": [
300
+ -1.9419843713000227,
301
+ 3.4050179211469533
302
+ ]
303
+ },
304
+ "behavior": {
305
+ "difference": 0.08333333333333215,
306
+ "ci95": [
307
+ -1.8518795120736862,
308
+ 2.1669680102528024
309
+ ]
310
+ }
311
+ },
312
+ "Deep8-Full": {
313
+ "color": {
314
+ "difference": 0.8333333333333321,
315
+ "ci95": [
316
+ -2.3809523809523796,
317
+ 4.051799085733517
318
+ ]
319
+ },
320
+ "count": {
321
+ "difference": -3.833333333333333,
322
+ "ci95": [
323
+ -7.741522152807696,
324
+ -0.17262390082673582
325
+ ]
326
+ },
327
+ "behavior": {
328
+ "difference": -1.5,
329
+ "ci95": [
330
+ -3.862156197590915,
331
+ 0.8220405373870728
332
+ ]
333
+ }
334
+ },
335
+ "Middle8-Full": {
336
+ "color": {
337
+ "difference": 2.833333333333332,
338
+ "ci95": [
339
+ 0.341793570219965,
340
+ 5.666666666666666
341
+ ]
342
+ },
343
+ "count": {
344
+ "difference": -0.8333333333333321,
345
+ "ci95": [
346
+ -4.477851128970534,
347
+ 2.4062436253510975
348
+ ]
349
+ },
350
+ "behavior": {
351
+ "difference": 1.0,
352
+ "ci95": [
353
+ -1.2043076726577198,
354
+ 3.107975547150466
355
+ ]
356
+ }
357
+ },
358
+ "Random8-Full": {
359
+ "color": {
360
+ "difference": 0.6166666666666654,
361
+ "ci95": [
362
+ -1.3499999999999996,
363
+ 3.0694619185675487
364
+ ]
365
+ },
366
+ "count": {
367
+ "difference": -0.4666666666666668,
368
+ "ci95": [
369
+ -3.333758503401361,
370
+ 1.96538709875784
371
+ ]
372
+ },
373
+ "behavior": {
374
+ "difference": 0.07499999999999929,
375
+ "ci95": [
376
+ -1.629657520345539,
377
+ 1.512899304749076
378
+ ]
379
+ }
380
+ },
381
+ "Shallow8-Full": {
382
+ "color": {
383
+ "difference": 1.5,
384
+ "ci95": [
385
+ -1.629649758454108,
386
+ 4.975124378109451
387
+ ]
388
+ },
389
+ "count": {
390
+ "difference": -1.833333333333333,
391
+ "ci95": [
392
+ -5.218952194758646,
393
+ 1.258161126085656
394
+ ]
395
+ },
396
+ "behavior": {
397
+ "difference": -0.16666666666666607,
398
+ "ci95": [
399
+ -2.3435467387311233,
400
+ 1.995625026660409
401
+ ]
402
+ }
403
+ },
404
+ "Top8-Full": {
405
+ "color": {
406
+ "difference": 1.166666666666666,
407
+ "ci95": [
408
+ -1.7242123756774457,
409
+ 4.355108877721944
410
+ ]
411
+ },
412
+ "count": {
413
+ "difference": -1.333333333333333,
414
+ "ci95": [
415
+ -4.431624851391792,
416
+ 1.738383309029046
417
+ ]
418
+ },
419
+ "behavior": {
420
+ "difference": -0.08333333333333393,
421
+ "ci95": [
422
+ -2.152799599172603,
423
+ 2.0720516281469554
424
+ ]
425
+ }
426
+ },
427
+ "TopDelta8-Full": {
428
+ "color": {
429
+ "difference": 2.5,
430
+ "ci95": [
431
+ 1.7763568394002505e-15,
432
+ 5.034870426829269
433
+ ]
434
+ },
435
+ "count": {
436
+ "difference": -0.6666666666666661,
437
+ "ci95": [
438
+ -3.780161604903872,
439
+ 2.3150839793281666
440
+ ]
441
+ },
442
+ "behavior": {
443
+ "difference": 0.9166666666666661,
444
+ "ci95": [
445
+ -0.898544170848408,
446
+ 2.7615857987030425
447
+ ]
448
+ }
449
+ }
450
+ }
451
+ }
reports/frozen_v3/fixed_internvl_replacement.json ADDED
@@ -0,0 +1,451 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "replacement",
3
+ "arms": {
4
+ "Deep8": {
5
+ "color": {
6
+ "gain": 5.5,
7
+ "ci95": [
8
+ 2.1975709475709477,
9
+ 9.153808195258971
10
+ ]
11
+ },
12
+ "count": {
13
+ "gain": 0.6666666666666666,
14
+ "ci95": [
15
+ -4.283120510774142,
16
+ 5.699597314196522
17
+ ]
18
+ },
19
+ "behavior": {
20
+ "gain": 3.0833333333333335,
21
+ "ci95": [
22
+ 0.041972000238318305,
23
+ 6.1282139502268365
24
+ ]
25
+ },
26
+ "runs": 3
27
+ },
28
+ "Full": {
29
+ "color": {
30
+ "gain": 5.833333333333333,
31
+ "ci95": [
32
+ 2.4560113519091846,
33
+ 9.479161358270058
34
+ ]
35
+ },
36
+ "count": {
37
+ "gain": 0.6666666666666666,
38
+ "ci95": [
39
+ -4.715958061219957,
40
+ 6.250560035842296
41
+ ]
42
+ },
43
+ "behavior": {
44
+ "gain": 3.25,
45
+ "ci95": [
46
+ 0.03493230625583557,
47
+ 6.558021532610946
48
+ ]
49
+ },
50
+ "runs": 3
51
+ },
52
+ "Middle8": {
53
+ "color": {
54
+ "gain": 6.166666666666667,
55
+ "ci95": [
56
+ 2.640046899426785,
57
+ 10.047846889952153
58
+ ]
59
+ },
60
+ "count": {
61
+ "gain": 0.0,
62
+ "ci95": [
63
+ -5.5555555555555545,
64
+ 5.641093474426809
65
+ ]
66
+ },
67
+ "behavior": {
68
+ "gain": 3.0833333333333335,
69
+ "ci95": [
70
+ -0.1524001016193441,
71
+ 6.519581833336135
72
+ ]
73
+ },
74
+ "runs": 3
75
+ },
76
+ "Random8": {
77
+ "color": {
78
+ "gain": 5.05,
79
+ "ci95": [
80
+ 1.8817071223251354,
81
+ 8.398106796116503
82
+ ]
83
+ },
84
+ "count": {
85
+ "gain": 0.8,
86
+ "ci95": [
87
+ -4.0308084772370485,
88
+ 5.868597155069721
89
+ ]
90
+ },
91
+ "behavior": {
92
+ "gain": 2.925,
93
+ "ci95": [
94
+ -0.028902785810680506,
95
+ 5.880289649718909
96
+ ]
97
+ },
98
+ "runs": 10
99
+ },
100
+ "Shallow8": {
101
+ "color": {
102
+ "gain": 4.5,
103
+ "ci95": [
104
+ 1.0415317357512954,
105
+ 8.243819975612459
106
+ ]
107
+ },
108
+ "count": {
109
+ "gain": 0.3333333333333333,
110
+ "ci95": [
111
+ -5.263854079643553,
112
+ 6.122896527031868
113
+ ]
114
+ },
115
+ "behavior": {
116
+ "gain": 2.4166666666666665,
117
+ "ci95": [
118
+ -0.9788146698554657,
119
+ 5.777629470307291
120
+ ]
121
+ },
122
+ "runs": 3
123
+ },
124
+ "Top8": {
125
+ "color": {
126
+ "gain": 4.833333333333333,
127
+ "ci95": [
128
+ 1.4771185835653606,
129
+ 8.421908429610614
130
+ ]
131
+ },
132
+ "count": {
133
+ "gain": -0.16666666666666666,
134
+ "ci95": [
135
+ -5.729916557591623,
136
+ 5.492762445887447
137
+ ]
138
+ },
139
+ "behavior": {
140
+ "gain": 2.3333333333333335,
141
+ "ci95": [
142
+ -0.916828529657477,
143
+ 5.568640197353072
144
+ ]
145
+ },
146
+ "runs": 3
147
+ },
148
+ "TopDelta8": {
149
+ "color": {
150
+ "gain": 6.0,
151
+ "ci95": [
152
+ 2.512562814070352,
153
+ 9.667063492063493
154
+ ]
155
+ },
156
+ "count": {
157
+ "gain": 0.3333333333333333,
158
+ "ci95": [
159
+ -4.957264957264957,
160
+ 5.582943662324023
161
+ ]
162
+ },
163
+ "behavior": {
164
+ "gain": 3.1666666666666665,
165
+ "ci95": [
166
+ -0.07104606737497907,
167
+ 6.437948987509699
168
+ ]
169
+ },
170
+ "runs": 3
171
+ }
172
+ },
173
+ "contrasts": {
174
+ "Top8-Random8": {
175
+ "color": {
176
+ "difference": -0.21666666666666679,
177
+ "ci95": [
178
+ -1.7863636363636373,
179
+ 1.2592563621193162
180
+ ]
181
+ },
182
+ "count": {
183
+ "difference": -0.9666666666666667,
184
+ "ci95": [
185
+ -3.583524557165861,
186
+ 1.7464901988953845
187
+ ]
188
+ },
189
+ "behavior": {
190
+ "difference": -0.5916666666666663,
191
+ "ci95": [
192
+ -2.124358758652281,
193
+ 0.9655486008411758
194
+ ]
195
+ }
196
+ },
197
+ "TopDelta8-Random8": {
198
+ "color": {
199
+ "difference": 0.9500000000000002,
200
+ "ci95": [
201
+ -2.7755575615619447e-18,
202
+ 2.349755536381937
203
+ ]
204
+ },
205
+ "count": {
206
+ "difference": -0.46666666666666673,
207
+ "ci95": [
208
+ -2.4317695104337864,
209
+ 1.4085655235762413
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "difference": 0.2416666666666667,
214
+ "ci95": [
215
+ -0.9317343049122525,
216
+ 1.551202012472187
217
+ ]
218
+ }
219
+ },
220
+ "Shallow8-Random8": {
221
+ "color": {
222
+ "difference": -0.5499999999999998,
223
+ "ci95": [
224
+ -2.4752475247524752,
225
+ 1.1836242138364783
226
+ ]
227
+ },
228
+ "count": {
229
+ "difference": -0.46666666666666673,
230
+ "ci95": [
231
+ -3.414695428361319,
232
+ 2.7127896613190745
233
+ ]
234
+ },
235
+ "behavior": {
236
+ "difference": -0.5083333333333333,
237
+ "ci95": [
238
+ -2.2997894154415555,
239
+ 1.3372060167433903
240
+ ]
241
+ }
242
+ },
243
+ "Middle8-Random8": {
244
+ "color": {
245
+ "difference": 1.1166666666666671,
246
+ "ci95": [
247
+ -0.338409475465312,
248
+ 2.947043066144653
249
+ ]
250
+ },
251
+ "count": {
252
+ "difference": -0.8,
253
+ "ci95": [
254
+ -3.4155755768732177,
255
+ 2.0574264394753117
256
+ ]
257
+ },
258
+ "behavior": {
259
+ "difference": 0.15833333333333366,
260
+ "ci95": [
261
+ -1.300025255127855,
262
+ 1.886065176539661
263
+ ]
264
+ }
265
+ },
266
+ "Deep8-Random8": {
267
+ "color": {
268
+ "difference": 0.4500000000000002,
269
+ "ci95": [
270
+ -0.6767266360782422,
271
+ 1.7058284667353703
272
+ ]
273
+ },
274
+ "count": {
275
+ "difference": -0.13333333333333341,
276
+ "ci95": [
277
+ -2.227644288819126,
278
+ 1.8501392804403283
279
+ ]
280
+ },
281
+ "behavior": {
282
+ "difference": 0.15833333333333366,
283
+ "ci95": [
284
+ -0.9475457822480527,
285
+ 1.2506931292087544
286
+ ]
287
+ }
288
+ },
289
+ "Top8-Shallow8": {
290
+ "color": {
291
+ "difference": 0.33333333333333304,
292
+ "ci95": [
293
+ -0.6472491909385107,
294
+ 1.4219065480565345
295
+ ]
296
+ },
297
+ "count": {
298
+ "difference": -0.5,
299
+ "ci95": [
300
+ -3.0895147095800852,
301
+ 1.8425460636515916
302
+ ]
303
+ },
304
+ "behavior": {
305
+ "difference": -0.08333333333333304,
306
+ "ci95": [
307
+ -1.4753264033196176,
308
+ 1.250011198196622
309
+ ]
310
+ }
311
+ },
312
+ "Deep8-Full": {
313
+ "color": {
314
+ "difference": -0.33333333333333304,
315
+ "ci95": [
316
+ -1.8518518518518512,
317
+ 0.9569377990430628
318
+ ]
319
+ },
320
+ "count": {
321
+ "difference": 0.0,
322
+ "ci95": [
323
+ -2.8709550757373354,
324
+ 2.4179620034542313
325
+ ]
326
+ },
327
+ "behavior": {
328
+ "difference": -0.16666666666666652,
329
+ "ci95": [
330
+ -1.8273203478127629,
331
+ 1.1345159142640158
332
+ ]
333
+ }
334
+ },
335
+ "Middle8-Full": {
336
+ "color": {
337
+ "difference": 0.3333333333333339,
338
+ "ci95": [
339
+ -1.2404549735564718,
340
+ 2.1166835528537673
341
+ ]
342
+ },
343
+ "count": {
344
+ "difference": -0.6666666666666666,
345
+ "ci95": [
346
+ -3.608450042144849,
347
+ 2.4563131967850667
348
+ ]
349
+ },
350
+ "behavior": {
351
+ "difference": -0.16666666666666652,
352
+ "ci95": [
353
+ -1.7780610883128138,
354
+ 1.6718349602314826
355
+ ]
356
+ }
357
+ },
358
+ "Random8-Full": {
359
+ "color": {
360
+ "difference": -0.7833333333333332,
361
+ "ci95": [
362
+ -2.1539682539682525,
363
+ 0.09393024815559997
364
+ ]
365
+ },
366
+ "count": {
367
+ "difference": 0.13333333333333341,
368
+ "ci95": [
369
+ -2.425782770584751,
370
+ 2.3811846689895484
371
+ ]
372
+ },
373
+ "behavior": {
374
+ "difference": -0.3250000000000002,
375
+ "ci95": [
376
+ -1.9248199954998866,
377
+ 0.8313538662677991
378
+ ]
379
+ }
380
+ },
381
+ "Shallow8-Full": {
382
+ "color": {
383
+ "difference": -1.333333333333333,
384
+ "ci95": [
385
+ -3.472304502369669,
386
+ 0.49757462686567183
387
+ ]
388
+ },
389
+ "count": {
390
+ "difference": -0.3333333333333333,
391
+ "ci95": [
392
+ -3.3874103552626122,
393
+ 2.966841186736475
394
+ ]
395
+ },
396
+ "behavior": {
397
+ "difference": -0.8333333333333335,
398
+ "ci95": [
399
+ -2.727315779882576,
400
+ 1.011495180903665
401
+ ]
402
+ }
403
+ },
404
+ "Top8-Full": {
405
+ "color": {
406
+ "difference": -1.0,
407
+ "ci95": [
408
+ -2.837047494201875,
409
+ 0.5025125628140703
410
+ ]
411
+ },
412
+ "count": {
413
+ "difference": -0.8333333333333333,
414
+ "ci95": [
415
+ -3.426791277258567,
416
+ 1.9906648078289892
417
+ ]
418
+ },
419
+ "behavior": {
420
+ "difference": -0.9166666666666665,
421
+ "ci95": [
422
+ -2.5295529286117837,
423
+ 0.7356618762855107
424
+ ]
425
+ }
426
+ },
427
+ "TopDelta8-Full": {
428
+ "color": {
429
+ "difference": 0.16666666666666696,
430
+ "ci95": [
431
+ -1.0002525252525252,
432
+ 1.4563327851168058
433
+ ]
434
+ },
435
+ "count": {
436
+ "difference": -0.3333333333333333,
437
+ "ci95": [
438
+ -2.724358974358974,
439
+ 1.8730125079499695
440
+ ]
441
+ },
442
+ "behavior": {
443
+ "difference": -0.08333333333333348,
444
+ "ci95": [
445
+ -1.5632915846466247,
446
+ 1.261240453313624
447
+ ]
448
+ }
449
+ }
450
+ }
451
+ }
reports/frozen_v3/joint_qwen_context.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "context",
3
+ "arms": {
4
+ "Joint": {
5
+ "color": {
6
+ "gain": -1.3333333333333333,
7
+ "ci95": [
8
+ -4.340466485507246,
9
+ 1.6026069518716577
10
+ ]
11
+ },
12
+ "count": {
13
+ "gain": -0.6666666666666666,
14
+ "ci95": [
15
+ -4.035087719298246,
16
+ 2.680269627708435
17
+ ]
18
+ },
19
+ "behavior": {
20
+ "gain": -1.0,
21
+ "ci95": [
22
+ -3.3041347979672726,
23
+ 1.2403811824635946
24
+ ]
25
+ },
26
+ "runs": 3
27
+ }
28
+ },
29
+ "contrasts": {}
30
+ }
reports/frozen_v3/joint_qwen_removal.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "removal",
3
+ "arms": {
4
+ "Joint": {
5
+ "color": {
6
+ "gain": 17.166666666666668,
7
+ "ci95": [
8
+ 10.76335406793049,
9
+ 23.880597014925375
10
+ ]
11
+ },
12
+ "count": {
13
+ "gain": 2.8333333333333335,
14
+ "ci95": [
15
+ -0.9524948735475052,
16
+ 6.666666666666667
17
+ ]
18
+ },
19
+ "behavior": {
20
+ "gain": 10.0,
21
+ "ci95": [
22
+ 6.417576413464781,
23
+ 13.700670858980695
24
+ ]
25
+ },
26
+ "runs": 3
27
+ }
28
+ },
29
+ "contrasts": {}
30
+ }
reports/frozen_v3/joint_qwen_replacement.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "replacement",
3
+ "arms": {
4
+ "Joint": {
5
+ "color": {
6
+ "gain": 1.3333333333333333,
7
+ "ci95": [
8
+ -2.0345884832806362,
9
+ 4.975124378109453
10
+ ]
11
+ },
12
+ "count": {
13
+ "gain": 0.3333333333333335,
14
+ "ci95": [
15
+ -4.568527918781726,
16
+ 4.912280701754386
17
+ ]
18
+ },
19
+ "behavior": {
20
+ "gain": 0.8333333333333334,
21
+ "ci95": [
22
+ -2.377475578750167,
23
+ 4.0000100010001
24
+ ]
25
+ },
26
+ "runs": 3
27
+ }
28
+ },
29
+ "contrasts": {}
30
+ }
reports/frozen_v3/per_run.csv ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ job,stage,backbone,behavior,name,items,effect,effect_ci_low,effect_ci_high,gain,gain_ci_low,gain_ci_high,original_accuracy,original_accuracy_change
2
+ eval_base_qwen_removal,base_qwen,qwen,removal,,400,84.25,80.72233040201004,87.75,,,,93.0,0.0
3
+ eval_base_qwen_replacement,base_qwen,qwen,replacement,,400,87.0,83.25,90.25,,,,92.5,0.0
4
+ eval_base_qwen_context,base_qwen,qwen,context,,400,4.25,2.0,6.75,,,,94.5,0.0
5
+ eval_base_internvl_removal,base_internvl,internvl,removal,,400,82.5,78.5,86.21733668341709,,,,86.5,0.0
6
+ eval_base_internvl_replacement,base_internvl,internvl,replacement,,400,76.75,72.25,81.00624999999991,,,,85.5,0.0
7
+ eval_base_internvl_context,base_internvl,internvl,context,,400,10.500000000000002,7.0,14.0,,,,85.75,0.0
8
+ eval_base_llava_removal,base_llava,llava,removal,,400,79.25,75.0,83.25,,,,80.5,0.0
9
+ eval_base_llava_replacement,base_llava,llava,replacement,,400,68.5,63.5,73.25,,,,81.0,0.0
10
+ eval_base_llava_context,base_llava,llava,context,,400,12.5,8.5,16.5,,,,81.25,0.0
11
+ eval_joint_qwen_removal_seed13,joint_qwen,qwen,removal,seed13,400,94.5,92.2248743718593,96.5298840413598,10.25,6.716141391257996,14.092783835066173,92.5,-0.5
12
+ eval_joint_qwen_replacement_seed13,joint_qwen,qwen,replacement,seed13,400,85.75,82.0,89.5,-1.25,-3.9616298120858704,1.462716553220801,91.75,-0.75
13
+ eval_joint_qwen_context_seed13,joint_qwen,qwen,context,seed13,400,5.0,2.5,7.5,-0.75,-3.3210029299968324,1.7657923269080076,93.0,-1.5
14
+ eval_joint_qwen_removal_seed42,joint_qwen,qwen,removal,seed42,400,93.75,91.25,96.0,9.5,5.810904591434834,13.563402684489551,93.5,0.5
15
+ eval_joint_qwen_replacement_seed42,joint_qwen,qwen,replacement,seed42,400,87.75,84.25,91.0,0.75,-1.9925597787365354,3.527467786470316,92.75,0.25
16
+ eval_joint_qwen_context_seed42,joint_qwen,qwen,context,seed42,400,5.75,3.0,8.75,-1.5,-4.135642969365116,1.1741413658384159,94.5,0.0
17
+ eval_joint_qwen_removal_seed73,joint_qwen,qwen,removal,seed73,400,94.5,92.2324120603015,96.51249815280035,10.25,6.548601313825358,14.128206171963207,94.25,1.25
18
+ eval_joint_qwen_replacement_seed73,joint_qwen,qwen,replacement,seed73,400,90.0,86.75,93.0,3.0,0.5755354773488343,5.488704005112152,94.25,1.75
19
+ eval_joint_qwen_context_seed73,joint_qwen,qwen,context,seed73,400,5.0,2.5,7.75,-0.75,-3.5000087508750877,1.976367082366181,95.25,0.75
20
+ eval_fixed_internvl_removal_Deep8_s13,fixed_internvl,internvl,removal,Deep8_s13,400,89.5,86.20851130653266,92.51741293532339,7.0,2.856394271724264,11.14869790030252,87.25,0.75
21
+ eval_fixed_internvl_removal_Deep8_s42,fixed_internvl,internvl,removal,Deep8_s42,400,90.25,87.0,93.25,7.75,3.7925674267756198,11.806977479423866,89.5,3.0
22
+ eval_fixed_internvl_removal_Deep8_s73,fixed_internvl,internvl,removal,Deep8_s73,400,91.0,88.0,93.75,8.5,4.648868058686165,12.491967307775496,89.0,2.5
23
+ eval_fixed_internvl_removal_Full_s13,fixed_internvl,internvl,removal,Full_s13,400,91.25,88.46984924623115,93.98492462311557,8.75,4.575139984625496,12.802678804001117,89.0,2.5
24
+ eval_fixed_internvl_removal_Full_s42,fixed_internvl,internvl,removal,Full_s42,400,92.25,89.51492537313433,94.75,9.75,6.186081068083378,13.499507283400813,89.75,3.25
25
+ eval_fixed_internvl_removal_Full_s73,fixed_internvl,internvl,removal,Full_s73,400,91.75,88.9748743718593,94.25,9.25,5.296997452268867,13.26356363472599,89.5,3.0
26
+ eval_fixed_internvl_removal_Middle8_s13,fixed_internvl,internvl,removal,Middle8_s13,400,91.25,88.27602611940299,94.0,8.75,4.758539822017734,12.832331881759965,89.25,2.75
27
+ eval_fixed_internvl_removal_Middle8_s42,fixed_internvl,internvl,removal,Middle8_s42,400,93.25,90.75,95.5087064676617,10.75,6.983985137210944,14.589239248202711,89.75,3.25
28
+ eval_fixed_internvl_removal_Middle8_s73,fixed_internvl,internvl,removal,Middle8_s73,400,93.75,91.25,96.0,11.25,7.487877505295552,15.296253163100983,87.5,1.0
29
+ eval_fixed_internvl_removal_Random8_8101_s13,fixed_internvl,internvl,removal,Random8_8101_s13,400,89.75,86.75,92.5776661942718,7.25,3.0288316786105303,11.506091143535162,89.75,3.25
30
+ eval_fixed_internvl_removal_Random8_8102_s13,fixed_internvl,internvl,removal,Random8_8102_s13,400,92.5,89.76990049751244,94.99120603015075,10.0,5.978529449954433,14.082705719142572,88.75,2.25
31
+ eval_fixed_internvl_removal_Random8_8103_s13,fixed_internvl,internvl,removal,Random8_8103_s13,400,92.5,89.72989949748744,95.0,10.0,6.25,13.852434534058393,89.0,2.5
32
+ eval_fixed_internvl_removal_Random8_8104_s13,fixed_internvl,internvl,removal,Random8_8104_s13,400,92.5,89.76990049751244,95.0,10.0,6.025566264156604,14.044537264646584,89.75,3.25
33
+ eval_fixed_internvl_removal_Random8_8105_s13,fixed_internvl,internvl,removal,Random8_8105_s13,400,92.5,89.8019801980198,95.0,10.0,6.047840015932582,14.109959960899976,88.75,2.25
34
+ eval_fixed_internvl_removal_Random8_8106_s13,fixed_internvl,internvl,removal,Random8_8106_s13,400,92.5,89.76492537313433,94.99246231155779,10.0,5.988591947427155,14.250996395079152,89.5,3.0
35
+ eval_fixed_internvl_removal_Random8_8107_s13,fixed_internvl,internvl,removal,Random8_8107_s13,400,93.5,91.0,95.75,11.0,7.002814054726367,15.225665147370142,89.5,3.0
36
+ eval_fixed_internvl_removal_Random8_8108_s13,fixed_internvl,internvl,removal,Random8_8108_s13,400,90.75,87.75,93.5,8.25,4.05340691090351,12.660350276169341,88.25,1.75
37
+ eval_fixed_internvl_removal_Random8_8109_s13,fixed_internvl,internvl,removal,Random8_8109_s13,400,90.5,87.5,93.25,8.0,3.8138417475486217,12.268682822167408,89.25,2.75
38
+ eval_fixed_internvl_removal_Random8_8110_s13,fixed_internvl,internvl,removal,Random8_8110_s13,400,91.25,88.26865671641791,94.0,8.75,4.499877809445236,13.07355787142443,88.5,2.0
39
+ eval_fixed_internvl_removal_Shallow8_s13,fixed_internvl,internvl,removal,Shallow8_s13,400,92.0,89.25,94.51741293532339,9.5,5.5442975776949925,14.03667832989211,89.5,3.0
40
+ eval_fixed_internvl_removal_Shallow8_s42,fixed_internvl,internvl,removal,Shallow8_s42,400,91.75,88.97738693467338,94.26243781094527,9.25,5.533456681351842,13.225616734977956,90.25,3.75
41
+ eval_fixed_internvl_removal_Shallow8_s73,fixed_internvl,internvl,removal,Shallow8_s73,400,91.0,88.0569306930693,93.75,8.5,4.449600872845292,12.967301043358336,89.25,2.75
42
+ eval_fixed_internvl_removal_Top8_s13,fixed_internvl,internvl,removal,Top8_s13,400,92.75,90.22107412060302,95.23492462311557,10.25,6.528353658536585,14.265732611940072,89.75,3.25
43
+ eval_fixed_internvl_removal_Top8_s42,fixed_internvl,internvl,removal,Top8_s42,400,91.0,88.2211055276382,93.7324120603015,8.5,4.486318152462668,12.738608896322752,90.5,4.0
44
+ eval_fixed_internvl_removal_Top8_s73,fixed_internvl,internvl,removal,Top8_s73,400,91.25,88.46730527638191,93.81338295541742,8.75,4.741181029525738,13.000124182675249,89.25,2.75
45
+ eval_fixed_internvl_removal_TopDelta8_s13,fixed_internvl,internvl,removal,TopDelta8_s13,400,92.0,89.25,94.50998134328357,9.5,5.743129854684139,13.505735618753315,89.0,2.5
46
+ eval_fixed_internvl_removal_TopDelta8_s42,fixed_internvl,internvl,removal,TopDelta8_s42,400,93.25,90.52114427860697,95.52964901477833,10.75,6.932901821658615,14.750077780008029,89.75,3.25
47
+ eval_fixed_internvl_removal_TopDelta8_s73,fixed_internvl,internvl,removal,TopDelta8_s73,400,92.75,90.22358668341708,95.2399811557789,10.25,6.289443561431274,14.321499116337728,88.75,2.25
48
+ eval_fixed_internvl_replacement_Deep8_s13,fixed_internvl,internvl,replacement,Deep8_s13,400,79.75,75.25,84.25,3.0,0.133932437254818,5.966764413012762,86.25,0.75
49
+ eval_fixed_internvl_replacement_Deep8_s42,fixed_internvl,internvl,replacement,Deep8_s42,400,79.5,74.75,84.0,2.75,-0.27166358070592905,5.814856205640149,87.0,1.5
50
+ eval_fixed_internvl_replacement_Deep8_s73,fixed_internvl,internvl,replacement,Deep8_s73,400,80.25,75.75,84.5,3.5,0.21596397288415434,6.811961074156188,86.5,1.0
51
+ eval_fixed_internvl_replacement_Full_s13,fixed_internvl,internvl,replacement,Full_s13,400,79.75,75.25,84.25,3.0,0.16002415458937186,6.007396339215746,87.25,1.75
52
+ eval_fixed_internvl_replacement_Full_s42,fixed_internvl,internvl,replacement,Full_s42,400,79.25,74.75,83.75,2.5,-0.682120380494702,5.678966191189281,87.5,2.0
53
+ eval_fixed_internvl_replacement_Full_s73,fixed_internvl,internvl,replacement,Full_s73,400,81.0,76.5,85.25,4.25,0.5520600570972053,7.843301249149277,88.0,2.5
54
+ eval_fixed_internvl_replacement_Middle8_s13,fixed_internvl,internvl,replacement,Middle8_s13,400,79.5,75.0,84.0,2.75,-0.4933711117838492,6.159362488011963,87.5,2.0
55
+ eval_fixed_internvl_replacement_Middle8_s42,fixed_internvl,internvl,replacement,Middle8_s42,400,79.25,74.5,83.75,2.5,-0.8948208936976589,5.915345637214831,88.0,2.5
56
+ eval_fixed_internvl_replacement_Middle8_s73,fixed_internvl,internvl,replacement,Middle8_s73,400,80.75,76.25,85.0,4.0,0.5252856234758084,7.503343451856497,85.5,0.0
57
+ eval_fixed_internvl_replacement_Random8_8101_s13,fixed_internvl,internvl,replacement,Random8_8101_s13,400,79.25,74.5,83.75,2.5,-0.5003200305925424,5.5222603421484235,87.25,1.75
58
+ eval_fixed_internvl_replacement_Random8_8102_s13,fixed_internvl,internvl,replacement,Random8_8102_s13,400,80.5,76.0,84.75,3.75,0.6678564195172717,7.026106981051789,88.0,2.5
59
+ eval_fixed_internvl_replacement_Random8_8103_s13,fixed_internvl,internvl,replacement,Random8_8103_s13,400,79.5,75.0,84.0,2.75,-0.2725918225075808,5.871895628963657,88.5,3.0
60
+ eval_fixed_internvl_replacement_Random8_8104_s13,fixed_internvl,internvl,replacement,Random8_8104_s13,400,79.25,74.75,83.5,2.5,-0.4765360205433707,5.488734366892677,87.75,2.25
61
+ eval_fixed_internvl_replacement_Random8_8105_s13,fixed_internvl,internvl,replacement,Random8_8105_s13,400,80.0,75.5,84.25,3.25,0.44065241228070173,6.246496006150154,87.25,1.75
62
+ eval_fixed_internvl_replacement_Random8_8106_s13,fixed_internvl,internvl,replacement,Random8_8106_s13,400,79.25,74.75,83.5,2.5,-0.5551212617207681,5.633482119804561,87.25,1.75
63
+ eval_fixed_internvl_replacement_Random8_8107_s13,fixed_internvl,internvl,replacement,Random8_8107_s13,400,79.0,74.25,83.5,2.25,-0.7073623602769836,5.404463830924286,87.5,2.0
64
+ eval_fixed_internvl_replacement_Random8_8108_s13,fixed_internvl,internvl,replacement,Random8_8108_s13,400,79.75,75.25,84.0,3.0,0.036945765917643515,6.020628959724555,87.75,2.25
65
+ eval_fixed_internvl_replacement_Random8_8109_s13,fixed_internvl,internvl,replacement,Random8_8109_s13,400,80.25,75.75,84.5,3.5,0.459598488093505,6.697603720279664,86.75,1.25
66
+ eval_fixed_internvl_replacement_Random8_8110_s13,fixed_internvl,internvl,replacement,Random8_8110_s13,400,80.0,75.5,84.25,3.25,0.26331783935237535,6.266427609511189,88.0,2.5
67
+ eval_fixed_internvl_replacement_Shallow8_s13,fixed_internvl,internvl,replacement,Shallow8_s13,400,78.5,73.75,83.0,1.75,-1.7095732140764122,5.023653976120748,88.0,2.5
68
+ eval_fixed_internvl_replacement_Shallow8_s42,fixed_internvl,internvl,replacement,Shallow8_s42,400,80.0,75.5,84.25,3.25,0.009750243756093808,6.529363759546685,89.25,3.75
69
+ eval_fixed_internvl_replacement_Shallow8_s73,fixed_internvl,internvl,replacement,Shallow8_s73,400,79.0,74.25,83.5,2.25,-1.4482842521146206,5.853062644592245,88.75,3.25
70
+ eval_fixed_internvl_replacement_Top8_s13,fixed_internvl,internvl,replacement,Top8_s13,400,79.0,74.25,83.5,2.25,-0.9980172478722602,5.4650243194773,88.25,2.75
71
+ eval_fixed_internvl_replacement_Top8_s42,fixed_internvl,internvl,replacement,Top8_s42,400,79.0,74.5,83.5,2.25,-1.0007784195991993,5.523755951769831,88.0,2.5
72
+ eval_fixed_internvl_replacement_Top8_s73,fixed_internvl,internvl,replacement,Top8_s73,400,79.25,74.75,83.5,2.5,-1.3759245675163518,6.256406410160254,88.25,2.75
73
+ eval_fixed_internvl_replacement_TopDelta8_s13,fixed_internvl,internvl,replacement,TopDelta8_s13,400,80.0,75.5,84.25,3.25,0.25,6.377942068147662,87.0,1.5
74
+ eval_fixed_internvl_replacement_TopDelta8_s42,fixed_internvl,internvl,replacement,TopDelta8_s42,400,79.25,74.75,83.5,2.5,-0.6737121434893049,5.702384412548396,87.0,1.5
75
+ eval_fixed_internvl_replacement_TopDelta8_s73,fixed_internvl,internvl,replacement,TopDelta8_s73,400,80.5,76.0,84.75,3.75,0.21452958898992525,7.231129720625427,85.5,0.0
76
+ eval_fixed_internvl_context_Deep8_s13,fixed_internvl,internvl,context,Deep8_s13,400,7.5,4.25,10.75,3.0,-0.7131418276448804,7.075470586785437,88.75,3.0
77
+ eval_fixed_internvl_context_Deep8_s42,fixed_internvl,internvl,context,Deep8_s42,400,7.0,4.0,10.0,3.5,-0.2066662065165166,7.290803645401823,89.25,3.5
78
+ eval_fixed_internvl_context_Deep8_s73,fixed_internvl,internvl,context,Deep8_s73,400,7.0,4.0,10.0,3.5,-0.21833370837083704,7.2317727578033,88.75,3.0
79
+ eval_fixed_internvl_context_Full_s13,fixed_internvl,internvl,context,Full_s13,400,7.75,4.5,11.0,2.7500000000000004,-1.0406721788152409,6.67928668143074,89.25,3.5
80
+ eval_fixed_internvl_context_Full_s42,fixed_internvl,internvl,context,Full_s42,400,8.5,5.5,11.5,2.0,-1.7282007910908843,5.979127853018159,88.25,2.5
81
+ eval_fixed_internvl_context_Full_s73,fixed_internvl,internvl,context,Full_s73,400,8.0,5.0,11.0,2.5000000000000004,-0.9088653625472229,5.961035267630861,89.5,3.75
82
+ eval_fixed_internvl_context_Middle8_s13,fixed_internvl,internvl,context,Middle8_s13,400,8.75,5.75,12.0,1.75,-2.000860440329978,5.530740081216952,88.0,2.25
83
+ eval_fixed_internvl_context_Middle8_s42,fixed_internvl,internvl,context,Middle8_s42,400,7.25,4.243750000000006,10.5,3.2500000000000004,-0.5414150445590994,7.105761845893446,88.25,2.5
84
+ eval_fixed_internvl_context_Middle8_s73,fixed_internvl,internvl,context,Middle8_s73,400,8.5,5.5,11.5,2.0,-1.4423076923076923,5.472696901963612,88.25,2.5
85
+ eval_fixed_internvl_context_Random8_8101_s13,fixed_internvl,internvl,context,Random8_8101_s13,400,7.75,4.5,11.0,2.75,-0.9486136157886709,6.63928263848861,89.25,3.5
86
+ eval_fixed_internvl_context_Random8_8102_s13,fixed_internvl,internvl,context,Random8_8102_s13,400,9.75,6.75,13.0,0.7499999999999998,-2.9709035529268997,4.589413505719946,89.0,3.25
87
+ eval_fixed_internvl_context_Random8_8103_s13,fixed_internvl,internvl,context,Random8_8103_s13,400,6.75,3.75,10.0,3.75,0.29672873898148794,7.322906018753209,89.0,3.25
88
+ eval_fixed_internvl_context_Random8_8104_s13,fixed_internvl,internvl,context,Random8_8104_s13,400,8.25,5.0,11.5,2.25,-1.6509536455818445,6.0358535853585344,87.75,2.0
89
+ eval_fixed_internvl_context_Random8_8105_s13,fixed_internvl,internvl,context,Random8_8105_s13,400,8.75,5.5,12.0,1.75,-1.778959435831456,5.553937392683311,89.5,3.75
90
+ eval_fixed_internvl_context_Random8_8106_s13,fixed_internvl,internvl,context,Random8_8106_s13,400,8.75,5.5,12.0,1.75,-1.8127907602510402,5.374044584050878,89.5,3.75
91
+ eval_fixed_internvl_context_Random8_8107_s13,fixed_internvl,internvl,context,Random8_8107_s13,400,8.25,5.0,11.5,2.25,-1.500015006002401,6.007646958304846,89.0,3.25
92
+ eval_fixed_internvl_context_Random8_8108_s13,fixed_internvl,internvl,context,Random8_8108_s13,400,8.75,5.75,11.75,1.75,-2.0680585394546007,5.776340150023978,89.0,3.25
93
+ eval_fixed_internvl_context_Random8_8109_s13,fixed_internvl,internvl,context,Random8_8109_s13,400,9.25,6.25,12.5,1.25,-2.395819113717355,5.0920234357023215,89.25,3.5
94
+ eval_fixed_internvl_context_Random8_8110_s13,fixed_internvl,internvl,context,Random8_8110_s13,400,8.0,5.0,11.0,2.5,-1.1211394271731712,6.25,89.25,3.5
95
+ eval_fixed_internvl_context_Shallow8_s13,fixed_internvl,internvl,context,Shallow8_s13,400,5.250000000000001,2.25,8.25,5.25,1.4605732782161769,9.17656249999999,89.75,4.0
96
+ eval_fixed_internvl_context_Shallow8_s42,fixed_internvl,internvl,context,Shallow8_s42,400,8.5,5.5,11.75,2.0,-1.7329261471820259,5.857598172119088,89.5,3.75
97
+ eval_fixed_internvl_context_Shallow8_s73,fixed_internvl,internvl,context,Shallow8_s73,400,8.25,5.25,11.5,2.25,-1.6798779725438222,6.500543478260867,88.0,2.25
98
+ eval_fixed_internvl_context_Top8_s13,fixed_internvl,internvl,context,Top8_s13,400,8.25,5.25,11.5,2.25,-1.5,6.023976550414375,87.75,2.0
99
+ eval_fixed_internvl_context_Top8_s42,fixed_internvl,internvl,context,Top8_s42,400,7.0,3.9937500000000057,10.25,3.5,-0.0453391646687402,7.250412640901769,88.25,2.5
100
+ eval_fixed_internvl_context_Top8_s73,fixed_internvl,internvl,context,Top8_s73,400,7.0,4.0,10.25,3.5,0.14966378964271373,7.179884223950387,88.25,2.5
101
+ eval_fixed_internvl_context_TopDelta8_s13,fixed_internvl,internvl,context,TopDelta8_s13,400,8.75,5.75,12.0,1.75,-1.7657555446420992,5.469531249999996,87.75,2.0
102
+ eval_fixed_internvl_context_TopDelta8_s42,fixed_internvl,internvl,context,TopDelta8_s42,400,7.25,4.0,10.5,3.2500000000000004,-0.5137555022008803,7.214180354508861,89.0,3.25
103
+ eval_fixed_internvl_context_TopDelta8_s73,fixed_internvl,internvl,context,TopDelta8_s73,400,7.75,4.75,10.75,2.75,-1.0637318305001233,6.754797149122804,89.25,3.5
104
+ eval_single_internvl_removal_L01_s13,single_internvl,internvl,removal,L01_s13,400,84.75,81.21476130653267,88.25,2.25,-0.8067265712179895,5.422825376773071,89.75,3.25
105
+ eval_single_internvl_removal_L02_s13,single_internvl,internvl,removal,L02_s13,400,84.75,81.21356783919597,88.0667025088496,2.25,-0.741377665288737,5.354181079244707,90.0,3.5
106
+ eval_single_internvl_removal_L07_s13,single_internvl,internvl,removal,L07_s13,400,89.25,86.0,92.25,6.75,2.7626606399757736,10.758627872651084,88.75,2.25
107
+ eval_single_internvl_removal_L09_s13,single_internvl,internvl,removal,L09_s13,400,88.0,84.75,91.01990049751244,5.5,1.6650991548451843,9.48074463899272,89.5,3.0
108
+ eval_single_internvl_removal_L11_s13,single_internvl,internvl,removal,L11_s13,400,89.75,86.71984924623115,92.54213560210701,7.25,3.7335642023222713,10.937554178224687,89.25,2.75
109
+ eval_single_internvl_removal_L16_s13,single_internvl,internvl,removal,L16_s13,400,89.75,86.5273631840796,92.51871890547264,7.25,3.473338104337174,11.41884378260093,89.5,3.0
110
+ eval_single_internvl_removal_L20_s13,single_internvl,internvl,removal,L20_s13,400,91.0,88.0,93.75,8.5,4.9918503381439425,12.470606629143212,89.5,3.0
111
+ eval_single_internvl_removal_L25_s13,single_internvl,internvl,removal,L25_s13,400,89.75,86.5,92.75,7.25,3.1573756909695843,11.548631821036997,88.75,2.25
112
+ eval_single_internvl_removal_L29_s13,single_internvl,internvl,removal,L29_s13,400,90.0,86.7798196517413,92.98492462311557,7.5,3.544871871150046,11.536953773471842,89.5,3.0
113
+ eval_single_internvl_removal_L34_s13,single_internvl,internvl,removal,L34_s13,400,88.75,85.46482412060301,91.75,6.25,2.6447498748085305,9.89256471681922,88.75,2.25
114
+ eval_single_internvl_replacement_L01_s13,single_internvl,internvl,replacement,L01_s13,400,77.5,73.0,82.0,0.7500000000000002,-1.904864949776407,3.3343535310393158,87.25,1.75
115
+ eval_single_internvl_replacement_L02_s13,single_internvl,internvl,replacement,L02_s13,400,77.5,72.75,82.0,0.75,-1.7565732484240026,3.340770753133655,87.25,1.75
116
+ eval_single_internvl_replacement_L07_s13,single_internvl,internvl,replacement,L07_s13,400,78.25,73.75,82.75,1.5,-1.2689268895210943,4.266941740931832,88.25,2.75
117
+ eval_single_internvl_replacement_L09_s13,single_internvl,internvl,replacement,L09_s13,400,78.5,73.99375,83.0,1.75,-1.115608798189774,4.709702006909681,88.0,2.5
118
+ eval_single_internvl_replacement_L11_s13,single_internvl,internvl,replacement,L11_s13,400,78.0,73.25,82.5,1.2500000000000002,-1.6821572564404461,4.2092315938910945,88.0,2.5
119
+ eval_single_internvl_replacement_L16_s13,single_internvl,internvl,replacement,L16_s13,400,80.0,75.5,84.25,3.25,0.07943779537896004,6.524006038701953,88.5,3.0
120
+ eval_single_internvl_replacement_L20_s13,single_internvl,internvl,replacement,L20_s13,400,80.0,75.5,84.25,3.25,-0.0009716893603148982,6.62662292928895,88.0,2.5
121
+ eval_single_internvl_replacement_L25_s13,single_internvl,internvl,replacement,L25_s13,400,79.25,74.75,83.75,2.5,-0.6082272139162945,5.660514010539885,86.75,1.25
122
+ eval_single_internvl_replacement_L29_s13,single_internvl,internvl,replacement,L29_s13,400,77.75,73.0,82.25,1.0000000000000002,-1.972781046881313,3.8965921868572906,87.5,2.0
123
+ eval_single_internvl_replacement_L34_s13,single_internvl,internvl,replacement,L34_s13,400,78.25,73.75,82.5,1.5000000000000002,-1.0660448389882267,4.095679921546328,87.0,1.5
124
+ eval_single_internvl_context_L01_s13,single_internvl,internvl,context,L01_s13,400,7.75,4.5,11.0,2.75,-0.9800188879127214,6.555780578057806,87.25,1.5
125
+ eval_single_internvl_context_L02_s13,single_internvl,internvl,context,L02_s13,400,7.25,4.0,10.5,3.2500000000000004,-0.11554094741067852,6.813306330633063,86.75,1.0
126
+ eval_single_internvl_context_L07_s13,single_internvl,internvl,context,L07_s13,400,8.0,4.5,11.5,2.5000000000000004,-1.2755678320738824,6.4819444444444425,88.5,2.75
127
+ eval_single_internvl_context_L09_s13,single_internvl,internvl,context,L09_s13,400,8.75,5.5,12.25,1.75,-1.7086297740999141,5.441599359855966,88.75,3.0
128
+ eval_single_internvl_context_L11_s13,single_internvl,internvl,context,L11_s13,400,9.5,6.0,13.0,1.0,-2.9426614935162325,5.126663050541317,88.0,2.25
129
+ eval_single_internvl_context_L16_s13,single_internvl,internvl,context,L16_s13,400,8.25,5.25,11.5,2.25,-1.1105623381637695,6.108861758763935,88.0,2.25
130
+ eval_single_internvl_context_L20_s13,single_internvl,internvl,context,L20_s13,400,9.0,5.75,12.25,1.5,-1.9800744163479962,5.112937480623073,88.75,3.0
131
+ eval_single_internvl_context_L25_s13,single_internvl,internvl,context,L25_s13,400,8.25,5.0,11.5,2.25,-1.2933704860056685,5.923184095302921,88.25,2.5
132
+ eval_single_internvl_context_L29_s13,single_internvl,internvl,context,L29_s13,400,9.25,6.25,12.5,1.25,-2.252752299343037,5.000559454785878,88.5,2.75
133
+ eval_single_internvl_context_L34_s13,single_internvl,internvl,context,L34_s13,400,7.75,4.25,11.006249999999909,2.7500000000000004,-1.1416853126282702,6.977125546464398,88.0,2.25
134
+ eval_single_llava_removal_L01_s13,single_llava,llava,removal,L01_s13,400,88.5,85.25,91.51868781094527,9.25,5.142881597138968,13.636871325705943,78.75,-1.75
135
+ eval_single_llava_removal_L02_s13,single_llava,llava,removal,L02_s13,400,85.5,81.94971733668342,88.78240049751244,6.25,1.6090638525344352,11.239593392046283,75.5,-5.0
136
+ eval_single_llava_removal_L06_s13,single_llava,llava,removal,L06_s13,400,89.5,86.22609924623114,92.5,10.25,6.217265689904091,14.493690890990543,80.0,-0.5
137
+ eval_single_llava_removal_L08_s13,single_llava,llava,removal,L08_s13,400,88.5,85.25,91.52611940298507,9.25,5.1586331095872895,13.585090843405434,79.25,-1.25
138
+ eval_single_llava_removal_L10_s13,single_llava,llava,removal,L10_s13,400,90.5,87.5310945273632,93.25,11.25,7.115203128628778,15.669871857280869,79.5,-1.0
139
+ eval_single_llava_removal_L14_s13,single_llava,llava,removal,L14_s13,400,89.75,86.46984924623115,92.75,10.5,6.452646768436243,14.983362424869926,81.0,0.5
140
+ eval_single_llava_removal_L18_s13,single_llava,llava,removal,L18_s13,400,88.0,84.53358208955223,91.23492462311557,8.75,4.36626958475698,13.041379126693098,82.5,2.0
141
+ eval_single_llava_removal_L22_s13,single_llava,llava,removal,L22_s13,400,89.75,86.71984924623115,92.73618090452261,10.5,6.497500148216218,14.925461449806292,81.5,1.0
142
+ eval_single_llava_removal_L26_s13,single_llava,llava,removal,L26_s13,400,85.75,82.20979899497488,89.03731435643564,6.5,2.6321138211382116,10.546899293754509,81.75,1.25
143
+ eval_single_llava_removal_L30_s13,single_llava,llava,removal,L30_s13,400,89.0,85.7211055276382,92.0,9.75,5.611576437951277,14.282649948085417,83.75,3.25
144
+ eval_single_llava_replacement_L01_s13,single_llava,llava,replacement,L01_s13,400,67.5,62.25,72.5,-1.0,-4.976266438879213,3.0546531505058674,77.0,-4.0
145
+ eval_single_llava_replacement_L02_s13,single_llava,llava,replacement,L02_s13,400,67.0,61.75,72.0,-1.5000000000000004,-5.3814031440538095,2.7080512007290194,74.75,-6.25
146
+ eval_single_llava_replacement_L06_s13,single_llava,llava,replacement,L06_s13,400,66.5,61.25,71.75,-2.0,-6.17859386804219,2.296907461640508,76.25,-4.75
147
+ eval_single_llava_replacement_L08_s13,single_llava,llava,replacement,L08_s13,400,65.0,59.75,70.25,-3.5,-7.5131492665535005,0.613791719025769,74.0,-7.0
148
+ eval_single_llava_replacement_L10_s13,single_llava,llava,replacement,L10_s13,400,69.0,64.0,74.0,0.5,-3.5706329606488394,4.712367166174262,77.5,-3.5
149
+ eval_single_llava_replacement_L14_s13,single_llava,llava,replacement,L14_s13,400,65.0,59.75,70.25,-3.5,-7.797335793357239,0.7975996096261714,75.75,-5.25
150
+ eval_single_llava_replacement_L18_s13,single_llava,llava,replacement,L18_s13,400,67.0,61.75,72.0,-1.5000000000000004,-5.48551157337036,2.680202128574467,75.75,-5.25
151
+ eval_single_llava_replacement_L22_s13,single_llava,llava,replacement,L22_s13,400,66.0,60.75,71.25,-2.5,-6.612527369599109,1.8048049812029971,73.5,-7.5
152
+ eval_single_llava_replacement_L26_s13,single_llava,llava,replacement,L26_s13,400,67.75,62.5,72.75,-0.75,-4.5458532695374805,3.291276580833568,76.75,-4.25
153
+ eval_single_llava_replacement_L30_s13,single_llava,llava,replacement,L30_s13,400,67.75,62.5,72.75,-0.7500000000000002,-3.9838911210774635,2.644496880499118,78.0,-3.0
154
+ eval_single_llava_context_L01_s13,single_llava,llava,context,L01_s13,400,10.5,7.0,14.0,2.0,-2.047191712236892,6.097149013672297,81.75,0.5
155
+ eval_single_llava_context_L02_s13,single_llava,llava,context,L02_s13,400,9.5,6.0,13.0,3.0,-1.022802887059845,7.169356197249638,81.75,0.5
156
+ eval_single_llava_context_L06_s13,single_llava,llava,context,L06_s13,400,9.5,6.0,13.0,3.0,-1.2488399525333325,7.152922077865719,81.25,0.0
157
+ eval_single_llava_context_L08_s13,single_llava,llava,context,L08_s13,400,8.75,5.25,12.25,3.75,-0.49279556650246303,8.069760879156807,82.25,1.0
158
+ eval_single_llava_context_L10_s13,single_llava,llava,context,L10_s13,400,12.25,9.0,15.75,0.25,-4.246823119811982,4.598743639655034,81.0,-0.25
159
+ eval_single_llava_context_L14_s13,single_llava,llava,context,L14_s13,400,9.5,6.0,13.0,3.0,-1.25,7.214662523777282,82.25,1.0
160
+ eval_single_llava_context_L18_s13,single_llava,llava,context,L18_s13,400,9.5,5.75,13.25,3.0,-1.1368897306397308,7.464407135382959,83.0,1.75
161
+ eval_single_llava_context_L22_s13,single_llava,llava,context,L22_s13,400,8.75,5.0,12.5,3.75,-0.4799292429242924,8.190972176996683,81.75,0.5
162
+ eval_single_llava_context_L26_s13,single_llava,llava,context,L26_s13,400,9.75,6.25,13.5,2.75,-1.1970661826842977,6.955142240616607,82.75,1.5
163
+ eval_single_llava_context_L30_s13,single_llava,llava,context,L30_s13,400,13.25,9.75,17.0,-0.75,-4.25,2.7513551953335282,84.25,3.0
reports/frozen_v3/per_run.json ADDED
The diff for this file is too large to render. See raw diff
 
reports/frozen_v3/single_internvl_context.json ADDED
@@ -0,0 +1,300 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "context",
3
+ "runs": [
4
+ {
5
+ "layer": 1,
6
+ "seed": 13,
7
+ "color": {
8
+ "items": 200,
9
+ "gain": -1.0,
10
+ "ci95": [
11
+ -4.285714285714286,
12
+ 2.0103040454799235
13
+ ]
14
+ },
15
+ "count": {
16
+ "items": 200,
17
+ "gain": 6.5,
18
+ "ci95": [
19
+ 0.0,
20
+ 13.461787564766839
21
+ ]
22
+ },
23
+ "behavior": {
24
+ "gain": 2.75,
25
+ "ci95": [
26
+ -0.9800188879127214,
27
+ 6.555780578057806
28
+ ]
29
+ }
30
+ },
31
+ {
32
+ "layer": 2,
33
+ "seed": 13,
34
+ "color": {
35
+ "items": 200,
36
+ "gain": -0.5,
37
+ "ci95": [
38
+ -3.0303030303030303,
39
+ 2.304414149172212
40
+ ]
41
+ },
42
+ "count": {
43
+ "items": 200,
44
+ "gain": 7.000000000000001,
45
+ "ci95": [
46
+ 0.529030910609858,
47
+ 13.588191699604737
48
+ ]
49
+ },
50
+ "behavior": {
51
+ "gain": 3.2500000000000004,
52
+ "ci95": [
53
+ -0.11554094741067852,
54
+ 6.813306330633063
55
+ ]
56
+ }
57
+ },
58
+ {
59
+ "layer": 7,
60
+ "seed": 13,
61
+ "color": {
62
+ "items": 200,
63
+ "gain": -2.0,
64
+ "ci95": [
65
+ -5.26379048582996,
66
+ 0.970873786407767
67
+ ]
68
+ },
69
+ "count": {
70
+ "items": 200,
71
+ "gain": 7.000000000000001,
72
+ "ci95": [
73
+ 0.0,
74
+ 14.57286432160804
75
+ ]
76
+ },
77
+ "behavior": {
78
+ "gain": 2.5000000000000004,
79
+ "ci95": [
80
+ -1.2755678320738824,
81
+ 6.4819444444444425
82
+ ]
83
+ }
84
+ },
85
+ {
86
+ "layer": 9,
87
+ "seed": 13,
88
+ "color": {
89
+ "items": 200,
90
+ "gain": -1.0,
91
+ "ci95": [
92
+ -4.306737670224512,
93
+ 2.0204583910167653
94
+ ]
95
+ },
96
+ "count": {
97
+ "items": 200,
98
+ "gain": 4.5,
99
+ "ci95": [
100
+ -1.7275461106655972,
101
+ 11.458956339712916
102
+ ]
103
+ },
104
+ "behavior": {
105
+ "gain": 1.75,
106
+ "ci95": [
107
+ -1.7086297740999141,
108
+ 5.441599359855966
109
+ ]
110
+ }
111
+ },
112
+ {
113
+ "layer": 11,
114
+ "seed": 13,
115
+ "color": {
116
+ "items": 200,
117
+ "gain": -2.0,
118
+ "ci95": [
119
+ -5.208938953488372,
120
+ 0.9526098901098888
121
+ ]
122
+ },
123
+ "count": {
124
+ "items": 200,
125
+ "gain": 4.0,
126
+ "ci95": [
127
+ -2.9275501622011295,
128
+ 11.822792520790296
129
+ ]
130
+ },
131
+ "behavior": {
132
+ "gain": 1.0,
133
+ "ci95": [
134
+ -2.9426614935162325,
135
+ 5.126663050541317
136
+ ]
137
+ }
138
+ },
139
+ {
140
+ "layer": 16,
141
+ "seed": 13,
142
+ "color": {
143
+ "items": 200,
144
+ "gain": -1.5,
145
+ "ci95": [
146
+ -4.7368421052631575,
147
+ 1.087105013067236
148
+ ]
149
+ },
150
+ "count": {
151
+ "items": 200,
152
+ "gain": 6.0,
153
+ "ci95": [
154
+ 0.0,
155
+ 13.147589239138524
156
+ ]
157
+ },
158
+ "behavior": {
159
+ "gain": 2.25,
160
+ "ci95": [
161
+ -1.1105623381637695,
162
+ 6.108861758763935
163
+ ]
164
+ }
165
+ },
166
+ {
167
+ "layer": 20,
168
+ "seed": 13,
169
+ "color": {
170
+ "items": 200,
171
+ "gain": -1.5,
172
+ "ci95": [
173
+ -4.7368421052631575,
174
+ 1.087105013067236
175
+ ]
176
+ },
177
+ "count": {
178
+ "items": 200,
179
+ "gain": 4.5,
180
+ "ci95": [
181
+ -1.5465920623898297,
182
+ 11.283505154639169
183
+ ]
184
+ },
185
+ "behavior": {
186
+ "gain": 1.5,
187
+ "ci95": [
188
+ -1.9800744163479962,
189
+ 5.112937480623073
190
+ ]
191
+ }
192
+ },
193
+ {
194
+ "layer": 25,
195
+ "seed": 13,
196
+ "color": {
197
+ "items": 200,
198
+ "gain": -1.0,
199
+ "ci95": [
200
+ -4.2867445054945055,
201
+ 2.000251256281406
202
+ ]
203
+ },
204
+ "count": {
205
+ "items": 200,
206
+ "gain": 5.5,
207
+ "ci95": [
208
+ -0.9902221565440126,
209
+ 12.386568902410033
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "gain": 2.25,
214
+ "ci95": [
215
+ -1.2933704860056685,
216
+ 5.923184095302921
217
+ ]
218
+ }
219
+ },
220
+ {
221
+ "layer": 29,
222
+ "seed": 13,
223
+ "color": {
224
+ "items": 200,
225
+ "gain": -0.5,
226
+ "ci95": [
227
+ -3.4482758620689653,
228
+ 1.9233091787439602
229
+ ]
230
+ },
231
+ "count": {
232
+ "items": 200,
233
+ "gain": 3.0,
234
+ "ci95": [
235
+ -3.8109040717736367,
236
+ 9.944823696161523
237
+ ]
238
+ },
239
+ "behavior": {
240
+ "gain": 1.25,
241
+ "ci95": [
242
+ -2.252752299343037,
243
+ 5.000559454785878
244
+ ]
245
+ }
246
+ },
247
+ {
248
+ "layer": 34,
249
+ "seed": 13,
250
+ "color": {
251
+ "items": 200,
252
+ "gain": -1.5,
253
+ "ci95": [
254
+ -5.377403464166534,
255
+ 2.5255729887709566
256
+ ]
257
+ },
258
+ "count": {
259
+ "items": 200,
260
+ "gain": 7.000000000000001,
261
+ "ci95": [
262
+ 0.0,
263
+ 14.436242618356504
264
+ ]
265
+ },
266
+ "behavior": {
267
+ "gain": 2.7500000000000004,
268
+ "ci95": [
269
+ -1.1416853126282702,
270
+ 6.977125546464398
271
+ ]
272
+ }
273
+ }
274
+ ],
275
+ "layers": [
276
+ 1,
277
+ 2,
278
+ 7,
279
+ 9,
280
+ 11,
281
+ 16,
282
+ 20,
283
+ 25,
284
+ 29,
285
+ 34
286
+ ],
287
+ "layer_gain": [
288
+ 2.75,
289
+ 3.2500000000000004,
290
+ 2.5000000000000004,
291
+ 1.75,
292
+ 1.0,
293
+ 2.25,
294
+ 1.5,
295
+ 2.25,
296
+ 1.25,
297
+ 2.7500000000000004
298
+ ],
299
+ "spearman_gamma_gain": -0.40729671426564323
300
+ }
reports/frozen_v3/single_internvl_removal.json ADDED
@@ -0,0 +1,300 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "removal",
3
+ "runs": [
4
+ {
5
+ "layer": 1,
6
+ "seed": 13,
7
+ "color": {
8
+ "items": 200,
9
+ "gain": -0.5,
10
+ "ci95": [
11
+ -4.371584699453552,
12
+ 3.5916230366492043
13
+ ]
14
+ },
15
+ "count": {
16
+ "items": 200,
17
+ "gain": 5.0,
18
+ "ci95": [
19
+ 0.0,
20
+ 9.95260663507109
21
+ ]
22
+ },
23
+ "behavior": {
24
+ "gain": 2.25,
25
+ "ci95": [
26
+ -0.8067265712179895,
27
+ 5.422825376773071
28
+ ]
29
+ }
30
+ },
31
+ {
32
+ "layer": 2,
33
+ "seed": 13,
34
+ "color": {
35
+ "items": 200,
36
+ "gain": 1.0,
37
+ "ci95": [
38
+ -2.689149560117302,
39
+ 5.025644276428009
40
+ ]
41
+ },
42
+ "count": {
43
+ "items": 200,
44
+ "gain": 3.5000000000000004,
45
+ "ci95": [
46
+ -1.0418030104712042,
47
+ 8.45087940578796
48
+ ]
49
+ },
50
+ "behavior": {
51
+ "gain": 2.25,
52
+ "ci95": [
53
+ -0.741377665288737,
54
+ 5.354181079244707
55
+ ]
56
+ }
57
+ },
58
+ {
59
+ "layer": 7,
60
+ "seed": 13,
61
+ "color": {
62
+ "items": 200,
63
+ "gain": 8.5,
64
+ "ci95": [
65
+ 2.2337160046657254,
66
+ 14.438811783252449
67
+ ]
68
+ },
69
+ "count": {
70
+ "items": 200,
71
+ "gain": 5.0,
72
+ "ci95": [
73
+ 0.0,
74
+ 10.309278350515465
75
+ ]
76
+ },
77
+ "behavior": {
78
+ "gain": 6.75,
79
+ "ci95": [
80
+ 2.7626606399757736,
81
+ 10.758627872651084
82
+ ]
83
+ }
84
+ },
85
+ {
86
+ "layer": 9,
87
+ "seed": 13,
88
+ "color": {
89
+ "items": 200,
90
+ "gain": 5.0,
91
+ "ci95": [
92
+ -1.0050251256281406,
93
+ 10.796186738519209
94
+ ]
95
+ },
96
+ "count": {
97
+ "items": 200,
98
+ "gain": 6.0,
99
+ "ci95": [
100
+ 1.0415317357512954,
101
+ 11.112403100775186
102
+ ]
103
+ },
104
+ "behavior": {
105
+ "gain": 5.5,
106
+ "ci95": [
107
+ 1.6650991548451843,
108
+ 9.48074463899272
109
+ ]
110
+ }
111
+ },
112
+ {
113
+ "layer": 11,
114
+ "seed": 13,
115
+ "color": {
116
+ "items": 200,
117
+ "gain": 9.0,
118
+ "ci95": [
119
+ 3.4995646766169153,
120
+ 14.438811783252449
121
+ ]
122
+ },
123
+ "count": {
124
+ "items": 200,
125
+ "gain": 5.5,
126
+ "ci95": [
127
+ 0.5460847575863272,
128
+ 10.312209898861568
129
+ ]
130
+ },
131
+ "behavior": {
132
+ "gain": 7.25,
133
+ "ci95": [
134
+ 3.7335642023222713,
135
+ 10.937554178224687
136
+ ]
137
+ }
138
+ },
139
+ {
140
+ "layer": 16,
141
+ "seed": 13,
142
+ "color": {
143
+ "items": 200,
144
+ "gain": 7.000000000000001,
145
+ "ci95": [
146
+ 0.9998756218905472,
147
+ 13.170731707317072
148
+ ]
149
+ },
150
+ "count": {
151
+ "items": 200,
152
+ "gain": 7.5,
153
+ "ci95": [
154
+ 2.55069667460893,
155
+ 12.821428571428568
156
+ ]
157
+ },
158
+ "behavior": {
159
+ "gain": 7.25,
160
+ "ci95": [
161
+ 3.473338104337174,
162
+ 11.41884378260093
163
+ ]
164
+ }
165
+ },
166
+ {
167
+ "layer": 20,
168
+ "seed": 13,
169
+ "color": {
170
+ "items": 200,
171
+ "gain": 8.5,
172
+ "ci95": [
173
+ 2.984339386809794,
174
+ 14.045579018688946
175
+ ]
176
+ },
177
+ "count": {
178
+ "items": 200,
179
+ "gain": 8.5,
180
+ "ci95": [
181
+ 3.979119182413058,
182
+ 13.334224598930478
183
+ ]
184
+ },
185
+ "behavior": {
186
+ "gain": 8.5,
187
+ "ci95": [
188
+ 4.9918503381439425,
189
+ 12.470606629143212
190
+ ]
191
+ }
192
+ },
193
+ {
194
+ "layer": 25,
195
+ "seed": 13,
196
+ "color": {
197
+ "items": 200,
198
+ "gain": 7.000000000000001,
199
+ "ci95": [
200
+ 1.075125064688632,
201
+ 13.298758865248221
202
+ ]
203
+ },
204
+ "count": {
205
+ "items": 200,
206
+ "gain": 7.5,
207
+ "ci95": [
208
+ 1.5147727272727274,
209
+ 13.021399456521737
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "gain": 7.25,
214
+ "ci95": [
215
+ 3.1573756909695843,
216
+ 11.548631821036997
217
+ ]
218
+ }
219
+ },
220
+ {
221
+ "layer": 29,
222
+ "seed": 13,
223
+ "color": {
224
+ "items": 200,
225
+ "gain": 7.5,
226
+ "ci95": [
227
+ 2.0613296865137807,
228
+ 13.208046321253866
229
+ ]
230
+ },
231
+ "count": {
232
+ "items": 200,
233
+ "gain": 7.5,
234
+ "ci95": [
235
+ 1.8957345971563981,
236
+ 12.82079646017699
237
+ ]
238
+ },
239
+ "behavior": {
240
+ "gain": 7.5,
241
+ "ci95": [
242
+ 3.544871871150046,
243
+ 11.536953773471842
244
+ ]
245
+ }
246
+ },
247
+ {
248
+ "layer": 34,
249
+ "seed": 13,
250
+ "color": {
251
+ "items": 200,
252
+ "gain": 9.0,
253
+ "ci95": [
254
+ 4.2105263157894735,
255
+ 14.079263508875151
256
+ ]
257
+ },
258
+ "count": {
259
+ "items": 200,
260
+ "gain": 3.5000000000000004,
261
+ "ci95": [
262
+ -1.5625,
263
+ 8.629441624365482
264
+ ]
265
+ },
266
+ "behavior": {
267
+ "gain": 6.25,
268
+ "ci95": [
269
+ 2.6447498748085305,
270
+ 9.89256471681922
271
+ ]
272
+ }
273
+ }
274
+ ],
275
+ "layers": [
276
+ 1,
277
+ 2,
278
+ 7,
279
+ 9,
280
+ 11,
281
+ 16,
282
+ 20,
283
+ 25,
284
+ 29,
285
+ 34
286
+ ],
287
+ "layer_gain": [
288
+ 2.25,
289
+ 2.25,
290
+ 6.75,
291
+ 5.5,
292
+ 7.25,
293
+ 7.25,
294
+ 8.5,
295
+ 7.25,
296
+ 7.5,
297
+ 6.25
298
+ ],
299
+ "spearman_gamma_gain": -0.5662208585049305
300
+ }
reports/frozen_v3/single_internvl_replacement.json ADDED
@@ -0,0 +1,300 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "replacement",
3
+ "runs": [
4
+ {
5
+ "layer": 1,
6
+ "seed": 13,
7
+ "color": {
8
+ "items": 200,
9
+ "gain": 3.5000000000000004,
10
+ "ci95": [
11
+ 0.9659524304833229,
12
+ 6.566489770804488
13
+ ]
14
+ },
15
+ "count": {
16
+ "items": 200,
17
+ "gain": -2.0,
18
+ "ci95": [
19
+ -6.418376034267166,
20
+ 2.0730814095434416
21
+ ]
22
+ },
23
+ "behavior": {
24
+ "gain": 0.7500000000000002,
25
+ "ci95": [
26
+ -1.904864949776407,
27
+ 3.3343535310393158
28
+ ]
29
+ }
30
+ },
31
+ {
32
+ "layer": 2,
33
+ "seed": 13,
34
+ "color": {
35
+ "items": 200,
36
+ "gain": 4.0,
37
+ "ci95": [
38
+ 1.0415317357512954,
39
+ 7.281553398058253
40
+ ]
41
+ },
42
+ "count": {
43
+ "items": 200,
44
+ "gain": -2.5,
45
+ "ci95": [
46
+ -6.372549019607843,
47
+ 1.5003787878787858
48
+ ]
49
+ },
50
+ "behavior": {
51
+ "gain": 0.75,
52
+ "ci95": [
53
+ -1.7565732484240026,
54
+ 3.340770753133655
55
+ ]
56
+ }
57
+ },
58
+ {
59
+ "layer": 7,
60
+ "seed": 13,
61
+ "color": {
62
+ "items": 200,
63
+ "gain": 4.0,
64
+ "ci95": [
65
+ 1.0415317357512954,
66
+ 7.281553398058253
67
+ ]
68
+ },
69
+ "count": {
70
+ "items": 200,
71
+ "gain": -1.0,
72
+ "ci95": [
73
+ -5.612244897959184,
74
+ 3.333732057416266
75
+ ]
76
+ },
77
+ "behavior": {
78
+ "gain": 1.5,
79
+ "ci95": [
80
+ -1.2689268895210943,
81
+ 4.266941740931832
82
+ ]
83
+ }
84
+ },
85
+ {
86
+ "layer": 9,
87
+ "seed": 13,
88
+ "color": {
89
+ "items": 200,
90
+ "gain": 4.5,
91
+ "ci95": [
92
+ 1.5075376884422111,
93
+ 8.001005025125624
94
+ ]
95
+ },
96
+ "count": {
97
+ "items": 200,
98
+ "gain": -1.0,
99
+ "ci95": [
100
+ -6.122448979591836,
101
+ 3.865676917638736
102
+ ]
103
+ },
104
+ "behavior": {
105
+ "gain": 1.75,
106
+ "ci95": [
107
+ -1.115608798189774,
108
+ 4.709702006909681
109
+ ]
110
+ }
111
+ },
112
+ {
113
+ "layer": 11,
114
+ "seed": 13,
115
+ "color": {
116
+ "items": 200,
117
+ "gain": 3.5000000000000004,
118
+ "ci95": [
119
+ 0.5492277992277993,
120
+ 6.63265306122449
121
+ ]
122
+ },
123
+ "count": {
124
+ "items": 200,
125
+ "gain": -1.0,
126
+ "ci95": [
127
+ -6.350050658561297,
128
+ 4.040404040404041
129
+ ]
130
+ },
131
+ "behavior": {
132
+ "gain": 1.2500000000000002,
133
+ "ci95": [
134
+ -1.6821572564404461,
135
+ 4.2092315938910945
136
+ ]
137
+ }
138
+ },
139
+ {
140
+ "layer": 16,
141
+ "seed": 13,
142
+ "color": {
143
+ "items": 200,
144
+ "gain": 6.0,
145
+ "ci95": [
146
+ 2.5122487437185934,
147
+ 9.743589743589743
148
+ ]
149
+ },
150
+ "count": {
151
+ "items": 200,
152
+ "gain": 0.5,
153
+ "ci95": [
154
+ -4.976374409360234,
155
+ 5.7976765585461205
156
+ ]
157
+ },
158
+ "behavior": {
159
+ "gain": 3.25,
160
+ "ci95": [
161
+ 0.07943779537896004,
162
+ 6.524006038701953
163
+ ]
164
+ }
165
+ },
166
+ {
167
+ "layer": 20,
168
+ "seed": 13,
169
+ "color": {
170
+ "items": 200,
171
+ "gain": 5.5,
172
+ "ci95": [
173
+ 2.116123642439432,
174
+ 9.090909090909092
175
+ ]
176
+ },
177
+ "count": {
178
+ "items": 200,
179
+ "gain": 1.0,
180
+ "ci95": [
181
+ -4.478171641791044,
182
+ 6.436444017536078
183
+ ]
184
+ },
185
+ "behavior": {
186
+ "gain": 3.25,
187
+ "ci95": [
188
+ -0.0009716893603148982,
189
+ 6.62662292928895
190
+ ]
191
+ }
192
+ },
193
+ {
194
+ "layer": 25,
195
+ "seed": 13,
196
+ "color": {
197
+ "items": 200,
198
+ "gain": 4.5,
199
+ "ci95": [
200
+ 1.4851485148514851,
201
+ 7.870655446086479
202
+ ]
203
+ },
204
+ "count": {
205
+ "items": 200,
206
+ "gain": 0.5,
207
+ "ci95": [
208
+ -4.976374409360234,
209
+ 5.7976765585461205
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "gain": 2.5,
214
+ "ci95": [
215
+ -0.6082272139162945,
216
+ 5.660514010539885
217
+ ]
218
+ }
219
+ },
220
+ {
221
+ "layer": 29,
222
+ "seed": 13,
223
+ "color": {
224
+ "items": 200,
225
+ "gain": 3.5000000000000004,
226
+ "ci95": [
227
+ 0.8766440131032127,
228
+ 6.567331478045754
229
+ ]
230
+ },
231
+ "count": {
232
+ "items": 200,
233
+ "gain": -1.5,
234
+ "ci95": [
235
+ -6.666666666666667,
236
+ 3.3828857630008455
237
+ ]
238
+ },
239
+ "behavior": {
240
+ "gain": 1.0000000000000002,
241
+ "ci95": [
242
+ -1.972781046881313,
243
+ 3.8965921868572906
244
+ ]
245
+ }
246
+ },
247
+ {
248
+ "layer": 34,
249
+ "seed": 13,
250
+ "color": {
251
+ "items": 200,
252
+ "gain": 3.5000000000000004,
253
+ "ci95": [
254
+ 0.8766440131032127,
255
+ 6.567331478045754
256
+ ]
257
+ },
258
+ "count": {
259
+ "items": 200,
260
+ "gain": -0.5,
261
+ "ci95": [
262
+ -4.787234042553192,
263
+ 3.684271161775406
264
+ ]
265
+ },
266
+ "behavior": {
267
+ "gain": 1.5000000000000002,
268
+ "ci95": [
269
+ -1.0660448389882267,
270
+ 4.095679921546328
271
+ ]
272
+ }
273
+ }
274
+ ],
275
+ "layers": [
276
+ 1,
277
+ 2,
278
+ 7,
279
+ 9,
280
+ 11,
281
+ 16,
282
+ 20,
283
+ 25,
284
+ 29,
285
+ 34
286
+ ],
287
+ "layer_gain": [
288
+ 0.7500000000000002,
289
+ 0.75,
290
+ 1.5,
291
+ 1.75,
292
+ 1.2500000000000002,
293
+ 3.25,
294
+ 3.25,
295
+ 2.5,
296
+ 1.0000000000000002,
297
+ 1.5000000000000002
298
+ ],
299
+ "spearman_gamma_gain": -0.4012176588288426
300
+ }
reports/frozen_v3/single_llava_context.json ADDED
@@ -0,0 +1,299 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "context",
3
+ "runs": [
4
+ {
5
+ "layer": 1,
6
+ "seed": 13,
7
+ "color": {
8
+ "items": 200,
9
+ "gain": 3.0,
10
+ "ci95": [
11
+ -1.442481884057971,
12
+ 7.575757575757576
13
+ ]
14
+ },
15
+ "count": {
16
+ "items": 200,
17
+ "gain": 1.0,
18
+ "ci95": [
19
+ -5.855100324402259,
20
+ 7.895042839657281
21
+ ]
22
+ },
23
+ "behavior": {
24
+ "gain": 2.0,
25
+ "ci95": [
26
+ -2.047191712236892,
27
+ 6.097149013672297
28
+ ]
29
+ }
30
+ },
31
+ {
32
+ "layer": 2,
33
+ "seed": 13,
34
+ "color": {
35
+ "items": 200,
36
+ "gain": 2.5,
37
+ "ci95": [
38
+ -2.128231525966598,
39
+ 7.3040325901523575
40
+ ]
41
+ },
42
+ "count": {
43
+ "items": 200,
44
+ "gain": 3.5000000000000004,
45
+ "ci95": [
46
+ -3.333732057416268,
47
+ 10.606060606060606
48
+ ]
49
+ },
50
+ "behavior": {
51
+ "gain": 3.0,
52
+ "ci95": [
53
+ -1.022802887059845,
54
+ 7.169356197249638
55
+ ]
56
+ }
57
+ },
58
+ {
59
+ "layer": 6,
60
+ "seed": 13,
61
+ "color": {
62
+ "items": 200,
63
+ "gain": 3.0,
64
+ "ci95": [
65
+ -0.9757293161166906,
66
+ 6.8631341425459045
67
+ ]
68
+ },
69
+ "count": {
70
+ "items": 200,
71
+ "gain": 3.0,
72
+ "ci95": [
73
+ -4.433497536945813,
74
+ 10.5277081592871
75
+ ]
76
+ },
77
+ "behavior": {
78
+ "gain": 3.0,
79
+ "ci95": [
80
+ -1.2488399525333325,
81
+ 7.152922077865719
82
+ ]
83
+ }
84
+ },
85
+ {
86
+ "layer": 8,
87
+ "seed": 13,
88
+ "color": {
89
+ "items": 200,
90
+ "gain": 2.0,
91
+ "ci95": [
92
+ -2.1052631578947367,
93
+ 6.315789473684211
94
+ ]
95
+ },
96
+ "count": {
97
+ "items": 200,
98
+ "gain": 5.5,
99
+ "ci95": [
100
+ -2.3006210165059646,
101
+ 12.935774495398652
102
+ ]
103
+ },
104
+ "behavior": {
105
+ "gain": 3.75,
106
+ "ci95": [
107
+ -0.49279556650246303,
108
+ 8.069760879156807
109
+ ]
110
+ }
111
+ },
112
+ {
113
+ "layer": 10,
114
+ "seed": 13,
115
+ "color": {
116
+ "items": 200,
117
+ "gain": -1.0,
118
+ "ci95": [
119
+ -5.853658536585366,
120
+ 3.773839877613461
121
+ ]
122
+ },
123
+ "count": {
124
+ "items": 200,
125
+ "gain": 1.5,
126
+ "ci95": [
127
+ -5.670103092783505,
128
+ 8.719072164948447
129
+ ]
130
+ },
131
+ "behavior": {
132
+ "gain": 0.25,
133
+ "ci95": [
134
+ -4.246823119811982,
135
+ 4.598743639655034
136
+ ]
137
+ }
138
+ },
139
+ {
140
+ "layer": 14,
141
+ "seed": 13,
142
+ "color": {
143
+ "items": 200,
144
+ "gain": 2.0,
145
+ "ci95": [
146
+ -3.1092130829015545,
147
+ 7.143772893772889
148
+ ]
149
+ },
150
+ "count": {
151
+ "items": 200,
152
+ "gain": 4.0,
153
+ "ci95": [
154
+ -2.5387218534426657,
155
+ 10.68047850208044
156
+ ]
157
+ },
158
+ "behavior": {
159
+ "gain": 3.0,
160
+ "ci95": [
161
+ -1.25,
162
+ 7.214662523777282
163
+ ]
164
+ }
165
+ },
166
+ {
167
+ "layer": 18,
168
+ "seed": 13,
169
+ "color": {
170
+ "items": 200,
171
+ "gain": 1.0,
172
+ "ci95": [
173
+ -4.1894034363247545,
174
+ 6.522012235088484
175
+ ]
176
+ },
177
+ "count": {
178
+ "items": 200,
179
+ "gain": 5.0,
180
+ "ci95": [
181
+ -1.4087844603164006,
182
+ 11.766154737757164
183
+ ]
184
+ },
185
+ "behavior": {
186
+ "gain": 3.0,
187
+ "ci95": [
188
+ -1.1368897306397308,
189
+ 7.464407135382959
190
+ ]
191
+ }
192
+ },
193
+ {
194
+ "layer": 22,
195
+ "seed": 13,
196
+ "color": {
197
+ "items": 200,
198
+ "gain": 1.0,
199
+ "ci95": [
200
+ -3.500439698492462,
201
+ 5.613448209472462
202
+ ]
203
+ },
204
+ "count": {
205
+ "items": 200,
206
+ "gain": 6.5,
207
+ "ci95": [
208
+ -0.5105379848274388,
209
+ 14.215686274509803
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "gain": 3.75,
214
+ "ci95": [
215
+ -0.4799292429242924,
216
+ 8.190972176996683
217
+ ]
218
+ }
219
+ },
220
+ {
221
+ "layer": 26,
222
+ "seed": 13,
223
+ "color": {
224
+ "items": 200,
225
+ "gain": 2.5,
226
+ "ci95": [
227
+ -1.9801980198019802,
228
+ 6.967101114147562
229
+ ]
230
+ },
231
+ "count": {
232
+ "items": 200,
233
+ "gain": 3.0,
234
+ "ci95": [
235
+ -3.535802184279342,
236
+ 9.853307163382588
237
+ ]
238
+ },
239
+ "behavior": {
240
+ "gain": 2.75,
241
+ "ci95": [
242
+ -1.1970661826842977,
243
+ 6.955142240616607
244
+ ]
245
+ }
246
+ },
247
+ {
248
+ "layer": 30,
249
+ "seed": 13,
250
+ "color": {
251
+ "items": 200,
252
+ "gain": -0.5,
253
+ "ci95": [
254
+ -4.592425431711146,
255
+ 3.5714285714285716
256
+ ]
257
+ },
258
+ "count": {
259
+ "items": 200,
260
+ "gain": -1.0,
261
+ "ci95": [
262
+ -6.599886929535448,
263
+ 4.712165365998222
264
+ ]
265
+ },
266
+ "behavior": {
267
+ "gain": -0.75,
268
+ "ci95": [
269
+ -4.25,
270
+ 2.7513551953335282
271
+ ]
272
+ }
273
+ }
274
+ ],
275
+ "layers": [
276
+ 1,
277
+ 2,
278
+ 6,
279
+ 8,
280
+ 10,
281
+ 14,
282
+ 18,
283
+ 22,
284
+ 26,
285
+ 30
286
+ ],
287
+ "layer_gain": [
288
+ 2.0,
289
+ 3.0,
290
+ 3.0,
291
+ 3.75,
292
+ 0.25,
293
+ 3.0,
294
+ 3.0,
295
+ 3.75,
296
+ 2.75,
297
+ -0.75
298
+ ]
299
+ }
reports/frozen_v3/single_llava_removal.json ADDED
@@ -0,0 +1,299 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "removal",
3
+ "runs": [
4
+ {
5
+ "layer": 1,
6
+ "seed": 13,
7
+ "color": {
8
+ "items": 200,
9
+ "gain": 13.0,
10
+ "ci95": [
11
+ 7.034296482412061,
12
+ 18.906769470184614
13
+ ]
14
+ },
15
+ "count": {
16
+ "items": 200,
17
+ "gain": 5.5,
18
+ "ci95": [
19
+ -0.5025760113699812,
20
+ 11.483672248803826
21
+ ]
22
+ },
23
+ "behavior": {
24
+ "gain": 9.25,
25
+ "ci95": [
26
+ 5.142881597138968,
27
+ 13.636871325705943
28
+ ]
29
+ }
30
+ },
31
+ {
32
+ "layer": 2,
33
+ "seed": 13,
34
+ "color": {
35
+ "items": 200,
36
+ "gain": 12.5,
37
+ "ci95": [
38
+ 5.6980204596784905,
39
+ 20.21276595744681
40
+ ]
41
+ },
42
+ "count": {
43
+ "items": 200,
44
+ "gain": 0.0,
45
+ "ci95": [
46
+ -6.167400881057269,
47
+ 6.598984771573604
48
+ ]
49
+ },
50
+ "behavior": {
51
+ "gain": 6.25,
52
+ "ci95": [
53
+ 1.6090638525344352,
54
+ 11.239593392046283
55
+ ]
56
+ }
57
+ },
58
+ {
59
+ "layer": 6,
60
+ "seed": 13,
61
+ "color": {
62
+ "items": 200,
63
+ "gain": 15.5,
64
+ "ci95": [
65
+ 9.895833333333334,
66
+ 21.579674689235645
67
+ ]
68
+ },
69
+ "count": {
70
+ "items": 200,
71
+ "gain": 5.0,
72
+ "ci95": [
73
+ -0.5559350333940493,
74
+ 11.000229357798164
75
+ ]
76
+ },
77
+ "behavior": {
78
+ "gain": 10.25,
79
+ "ci95": [
80
+ 6.217265689904091,
81
+ 14.493690890990543
82
+ ]
83
+ }
84
+ },
85
+ {
86
+ "layer": 8,
87
+ "seed": 13,
88
+ "color": {
89
+ "items": 200,
90
+ "gain": 16.5,
91
+ "ci95": [
92
+ 10.144927536231885,
93
+ 23.189264372553055
94
+ ]
95
+ },
96
+ "count": {
97
+ "items": 200,
98
+ "gain": 2.0,
99
+ "ci95": [
100
+ -3.175146771037182,
101
+ 7.620879120879111
102
+ ]
103
+ },
104
+ "behavior": {
105
+ "gain": 9.25,
106
+ "ci95": [
107
+ 5.1586331095872895,
108
+ 13.585090843405434
109
+ ]
110
+ }
111
+ },
112
+ {
113
+ "layer": 10,
114
+ "seed": 13,
115
+ "color": {
116
+ "items": 200,
117
+ "gain": 18.0,
118
+ "ci95": [
119
+ 11.650485436893204,
120
+ 24.726301735647525
121
+ ]
122
+ },
123
+ "count": {
124
+ "items": 200,
125
+ "gain": 4.5,
126
+ "ci95": [
127
+ -0.5524861878453039,
128
+ 10.054112554112546
129
+ ]
130
+ },
131
+ "behavior": {
132
+ "gain": 11.25,
133
+ "ci95": [
134
+ 7.115203128628778,
135
+ 15.669871857280869
136
+ ]
137
+ }
138
+ },
139
+ {
140
+ "layer": 14,
141
+ "seed": 13,
142
+ "color": {
143
+ "items": 200,
144
+ "gain": 16.5,
145
+ "ci95": [
146
+ 10.76829268292683,
147
+ 22.748815165876778
148
+ ]
149
+ },
150
+ "count": {
151
+ "items": 200,
152
+ "gain": 4.5,
153
+ "ci95": [
154
+ -1.5001884422110552,
155
+ 10.606060606060606
156
+ ]
157
+ },
158
+ "behavior": {
159
+ "gain": 10.5,
160
+ "ci95": [
161
+ 6.452646768436243,
162
+ 14.983362424869926
163
+ ]
164
+ }
165
+ },
166
+ {
167
+ "layer": 18,
168
+ "seed": 13,
169
+ "color": {
170
+ "items": 200,
171
+ "gain": 13.0,
172
+ "ci95": [
173
+ 6.498790322580645,
174
+ 19.474937343358388
175
+ ]
176
+ },
177
+ "count": {
178
+ "items": 200,
179
+ "gain": 4.5,
180
+ "ci95": [
181
+ -1.5001884422110552,
182
+ 10.550516343183808
183
+ ]
184
+ },
185
+ "behavior": {
186
+ "gain": 8.75,
187
+ "ci95": [
188
+ 4.36626958475698,
189
+ 13.041379126693098
190
+ ]
191
+ }
192
+ },
193
+ {
194
+ "layer": 22,
195
+ "seed": 13,
196
+ "color": {
197
+ "items": 200,
198
+ "gain": 15.0,
199
+ "ci95": [
200
+ 8.808290155440414,
201
+ 21.634615384615383
202
+ ]
203
+ },
204
+ "count": {
205
+ "items": 200,
206
+ "gain": 6.0,
207
+ "ci95": [
208
+ 0.5126903553299492,
209
+ 11.558699293569292
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "gain": 10.5,
214
+ "ci95": [
215
+ 6.497500148216218,
216
+ 14.925461449806292
217
+ ]
218
+ }
219
+ },
220
+ {
221
+ "layer": 26,
222
+ "seed": 13,
223
+ "color": {
224
+ "items": 200,
225
+ "gain": 11.0,
226
+ "ci95": [
227
+ 5.290941697191697,
228
+ 17.18836805555555
229
+ ]
230
+ },
231
+ "count": {
232
+ "items": 200,
233
+ "gain": 2.0,
234
+ "ci95": [
235
+ -3.5175879396984926,
236
+ 7.368939590355195
237
+ ]
238
+ },
239
+ "behavior": {
240
+ "gain": 6.5,
241
+ "ci95": [
242
+ 2.6321138211382116,
243
+ 10.546899293754509
244
+ ]
245
+ }
246
+ },
247
+ {
248
+ "layer": 30,
249
+ "seed": 13,
250
+ "color": {
251
+ "items": 200,
252
+ "gain": 15.0,
253
+ "ci95": [
254
+ 8.910631900886424,
255
+ 21.31997738590709
256
+ ]
257
+ },
258
+ "count": {
259
+ "items": 200,
260
+ "gain": 4.5,
261
+ "ci95": [
262
+ -0.975609756097561,
263
+ 10.153331526292439
264
+ ]
265
+ },
266
+ "behavior": {
267
+ "gain": 9.75,
268
+ "ci95": [
269
+ 5.611576437951277,
270
+ 14.282649948085417
271
+ ]
272
+ }
273
+ }
274
+ ],
275
+ "layers": [
276
+ 1,
277
+ 2,
278
+ 6,
279
+ 8,
280
+ 10,
281
+ 14,
282
+ 18,
283
+ 22,
284
+ 26,
285
+ 30
286
+ ],
287
+ "layer_gain": [
288
+ 9.25,
289
+ 6.25,
290
+ 10.25,
291
+ 9.25,
292
+ 11.25,
293
+ 10.5,
294
+ 8.75,
295
+ 10.5,
296
+ 6.5,
297
+ 9.75
298
+ ]
299
+ }
reports/frozen_v3/single_llava_replacement.json ADDED
@@ -0,0 +1,299 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "behavior": "replacement",
3
+ "runs": [
4
+ {
5
+ "layer": 1,
6
+ "seed": 13,
7
+ "color": {
8
+ "items": 200,
9
+ "gain": 4.0,
10
+ "ci95": [
11
+ 0.5050505050505051,
12
+ 7.651324085750309
13
+ ]
14
+ },
15
+ "count": {
16
+ "items": 200,
17
+ "gain": -6.0,
18
+ "ci95": [
19
+ -12.935774495398656,
20
+ 1.5469989744696897
21
+ ]
22
+ },
23
+ "behavior": {
24
+ "gain": -1.0,
25
+ "ci95": [
26
+ -4.976266438879213,
27
+ 3.0546531505058674
28
+ ]
29
+ }
30
+ },
31
+ {
32
+ "layer": 2,
33
+ "seed": 13,
34
+ "color": {
35
+ "items": 200,
36
+ "gain": 4.0,
37
+ "ci95": [
38
+ 0.48303372352285395,
39
+ 7.870655446086479
40
+ ]
41
+ },
42
+ "count": {
43
+ "items": 200,
44
+ "gain": -7.000000000000001,
45
+ "ci95": [
46
+ -14.354189669979142,
47
+ 0.4904411764705869
48
+ ]
49
+ },
50
+ "behavior": {
51
+ "gain": -1.5000000000000004,
52
+ "ci95": [
53
+ -5.3814031440538095,
54
+ 2.7080512007290194
55
+ ]
56
+ }
57
+ },
58
+ {
59
+ "layer": 6,
60
+ "seed": 13,
61
+ "color": {
62
+ "items": 200,
63
+ "gain": 4.0,
64
+ "ci95": [
65
+ 0.4431818181818188,
66
+ 7.805068597560974
67
+ ]
68
+ },
69
+ "count": {
70
+ "items": 200,
71
+ "gain": -8.0,
72
+ "ci95": [
73
+ -15.657821672219576,
74
+ 0.0
75
+ ]
76
+ },
77
+ "behavior": {
78
+ "gain": -2.0,
79
+ "ci95": [
80
+ -6.17859386804219,
81
+ 2.296907461640508
82
+ ]
83
+ }
84
+ },
85
+ {
86
+ "layer": 8,
87
+ "seed": 13,
88
+ "color": {
89
+ "items": 200,
90
+ "gain": 2.5,
91
+ "ci95": [
92
+ -1.442481884057971,
93
+ 6.467661691542289
94
+ ]
95
+ },
96
+ "count": {
97
+ "items": 200,
98
+ "gain": -9.5,
99
+ "ci95": [
100
+ -16.582469775474955,
101
+ -2.0405573396871453
102
+ ]
103
+ },
104
+ "behavior": {
105
+ "gain": -3.5,
106
+ "ci95": [
107
+ -7.5131492665535005,
108
+ 0.613791719025769
109
+ ]
110
+ }
111
+ },
112
+ {
113
+ "layer": 10,
114
+ "seed": 13,
115
+ "color": {
116
+ "items": 200,
117
+ "gain": 5.0,
118
+ "ci95": [
119
+ 1.0927476835352816,
120
+ 9.18381136238279
121
+ ]
122
+ },
123
+ "count": {
124
+ "items": 200,
125
+ "gain": -4.0,
126
+ "ci95": [
127
+ -11.220714531088973,
128
+ 3.5723490427098628
129
+ ]
130
+ },
131
+ "behavior": {
132
+ "gain": 0.5,
133
+ "ci95": [
134
+ -3.5706329606488394,
135
+ 4.712367166174262
136
+ ]
137
+ }
138
+ },
139
+ {
140
+ "layer": 14,
141
+ "seed": 13,
142
+ "color": {
143
+ "items": 200,
144
+ "gain": 3.0,
145
+ "ci95": [
146
+ -0.576704545454545,
147
+ 7.027027027027027
148
+ ]
149
+ },
150
+ "count": {
151
+ "items": 200,
152
+ "gain": -10.0,
153
+ "ci95": [
154
+ -17.55477620989968,
155
+ -2.1273781380164376
156
+ ]
157
+ },
158
+ "behavior": {
159
+ "gain": -3.5,
160
+ "ci95": [
161
+ -7.797335793357239,
162
+ 0.7975996096261714
163
+ ]
164
+ }
165
+ },
166
+ {
167
+ "layer": 18,
168
+ "seed": 13,
169
+ "color": {
170
+ "items": 200,
171
+ "gain": 4.0,
172
+ "ci95": [
173
+ 0.0,
174
+ 8.45853280765981
175
+ ]
176
+ },
177
+ "count": {
178
+ "items": 200,
179
+ "gain": -7.000000000000001,
180
+ "ci95": [
181
+ -14.136257866624359,
182
+ 0.5027048507845339
183
+ ]
184
+ },
185
+ "behavior": {
186
+ "gain": -1.5000000000000004,
187
+ "ci95": [
188
+ -5.48551157337036,
189
+ 2.680202128574467
190
+ ]
191
+ }
192
+ },
193
+ {
194
+ "layer": 22,
195
+ "seed": 13,
196
+ "color": {
197
+ "items": 200,
198
+ "gain": 1.5,
199
+ "ci95": [
200
+ -1.6044937036398137,
201
+ 4.878048780487805
202
+ ]
203
+ },
204
+ "count": {
205
+ "items": 200,
206
+ "gain": -6.5,
207
+ "ci95": [
208
+ -14.217436974789916,
209
+ 1.4564882784750168
210
+ ]
211
+ },
212
+ "behavior": {
213
+ "gain": -2.5,
214
+ "ci95": [
215
+ -6.612527369599109,
216
+ 1.8048049812029971
217
+ ]
218
+ }
219
+ },
220
+ {
221
+ "layer": 26,
222
+ "seed": 13,
223
+ "color": {
224
+ "items": 200,
225
+ "gain": 3.0,
226
+ "ci95": [
227
+ 0.0,
228
+ 6.011409506549136
229
+ ]
230
+ },
231
+ "count": {
232
+ "items": 200,
233
+ "gain": -4.5,
234
+ "ci95": [
235
+ -11.708751793400287,
236
+ 3.256578229064304
237
+ ]
238
+ },
239
+ "behavior": {
240
+ "gain": -0.75,
241
+ "ci95": [
242
+ -4.5458532695374805,
243
+ 3.291276580833568
244
+ ]
245
+ }
246
+ },
247
+ {
248
+ "layer": 30,
249
+ "seed": 13,
250
+ "color": {
251
+ "items": 200,
252
+ "gain": 2.0,
253
+ "ci95": [
254
+ -0.9952748818720468,
255
+ 5.208333333333333
256
+ ]
257
+ },
258
+ "count": {
259
+ "items": 200,
260
+ "gain": -3.5000000000000004,
261
+ "ci95": [
262
+ -9.43443161550737,
263
+ 2.592044301888854
264
+ ]
265
+ },
266
+ "behavior": {
267
+ "gain": -0.7500000000000002,
268
+ "ci95": [
269
+ -3.9838911210774635,
270
+ 2.644496880499118
271
+ ]
272
+ }
273
+ }
274
+ ],
275
+ "layers": [
276
+ 1,
277
+ 2,
278
+ 6,
279
+ 8,
280
+ 10,
281
+ 14,
282
+ 18,
283
+ 22,
284
+ 26,
285
+ 30
286
+ ],
287
+ "layer_gain": [
288
+ -1.0,
289
+ -1.5000000000000004,
290
+ -2.0,
291
+ -3.5,
292
+ 0.5,
293
+ -3.5,
294
+ -1.5000000000000004,
295
+ -2.5,
296
+ -0.75,
297
+ -0.7500000000000002
298
+ ]
299
+ }
reports/frozen_v3/summary.json ADDED
@@ -0,0 +1,3280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "complete": true,
3
+ "scope": "frozen v3 test evaluation of every formal adapter and matched base",
4
+ "protocol": {
5
+ "test_revision": "ba161c6fbb2829631ac9fef67ea296f0c4c1fbc4",
6
+ "test_items_per_behavior": 400,
7
+ "tasks": {
8
+ "color": 200,
9
+ "count": 200
10
+ },
11
+ "conditions": [
12
+ "original",
13
+ "treatment",
14
+ "control"
15
+ ],
16
+ "generation": "greedy, max_new_tokens=12, original alone; treatment/control paired",
17
+ "dtype": "float16",
18
+ "preprocessing": "original backbone-specific infer.py protocol",
19
+ "runtime_fix": "cuDNN SDPA disabled for all evaluation models per updated GitHub",
20
+ "evaluation_source_commit": "f0b8f2b91e412311052bbf92b674231df56d285e",
21
+ "training_source_commit": "52dd2b32d33fce9d26ae143868bc28f78c01d84b",
22
+ "training_parameters_frozen": true,
23
+ "test_set_used_for_parameter_selection": false,
24
+ "uploads_required": false,
25
+ "primary_metrics": "matched effects and paired gains, Color/Count equally weighted",
26
+ "secondary_metrics": "original-image accuracy, top-answer share, abstention",
27
+ "ci_resamples": {
28
+ "matched_effects": 10000,
29
+ "single_layer_gain": 2000,
30
+ "placement_and_joint": 3000
31
+ }
32
+ },
33
+ "validated_inference_jobs": 162,
34
+ "validated_records": 64800,
35
+ "formal_adapters": 147,
36
+ "analyses": {
37
+ "joint_qwen_removal": {
38
+ "behavior": "removal",
39
+ "arms": {
40
+ "Joint": {
41
+ "color": {
42
+ "gain": 17.166666666666668,
43
+ "ci95": [
44
+ 10.76335406793049,
45
+ 23.880597014925375
46
+ ]
47
+ },
48
+ "count": {
49
+ "gain": 2.8333333333333335,
50
+ "ci95": [
51
+ -0.9524948735475052,
52
+ 6.666666666666667
53
+ ]
54
+ },
55
+ "behavior": {
56
+ "gain": 10.0,
57
+ "ci95": [
58
+ 6.417576413464781,
59
+ 13.700670858980695
60
+ ]
61
+ },
62
+ "runs": 3
63
+ }
64
+ },
65
+ "contrasts": {}
66
+ },
67
+ "joint_qwen_replacement": {
68
+ "behavior": "replacement",
69
+ "arms": {
70
+ "Joint": {
71
+ "color": {
72
+ "gain": 1.3333333333333333,
73
+ "ci95": [
74
+ -2.0345884832806362,
75
+ 4.975124378109453
76
+ ]
77
+ },
78
+ "count": {
79
+ "gain": 0.3333333333333335,
80
+ "ci95": [
81
+ -4.568527918781726,
82
+ 4.912280701754386
83
+ ]
84
+ },
85
+ "behavior": {
86
+ "gain": 0.8333333333333334,
87
+ "ci95": [
88
+ -2.377475578750167,
89
+ 4.0000100010001
90
+ ]
91
+ },
92
+ "runs": 3
93
+ }
94
+ },
95
+ "contrasts": {}
96
+ },
97
+ "joint_qwen_context": {
98
+ "behavior": "context",
99
+ "arms": {
100
+ "Joint": {
101
+ "color": {
102
+ "gain": -1.3333333333333333,
103
+ "ci95": [
104
+ -4.340466485507246,
105
+ 1.6026069518716577
106
+ ]
107
+ },
108
+ "count": {
109
+ "gain": -0.6666666666666666,
110
+ "ci95": [
111
+ -4.035087719298246,
112
+ 2.680269627708435
113
+ ]
114
+ },
115
+ "behavior": {
116
+ "gain": -1.0,
117
+ "ci95": [
118
+ -3.3041347979672726,
119
+ 1.2403811824635946
120
+ ]
121
+ },
122
+ "runs": 3
123
+ }
124
+ },
125
+ "contrasts": {}
126
+ },
127
+ "fixed_internvl_removal": {
128
+ "behavior": "removal",
129
+ "arms": {
130
+ "Deep8": {
131
+ "color": {
132
+ "gain": 10.166666666666666,
133
+ "ci95": [
134
+ 4.537149096266674,
135
+ 15.781662433388119
136
+ ]
137
+ },
138
+ "count": {
139
+ "gain": 5.333333333333333,
140
+ "ci95": [
141
+ 0.0,
142
+ 10.72728106143948
143
+ ]
144
+ },
145
+ "behavior": {
146
+ "gain": 7.75,
147
+ "ci95": [
148
+ 3.916349296128708,
149
+ 11.591823277769754
150
+ ]
151
+ },
152
+ "runs": 3
153
+ },
154
+ "Full": {
155
+ "color": {
156
+ "gain": 9.333333333333334,
157
+ "ci95": [
158
+ 3.7233612201335817,
159
+ 15.183566882250283
160
+ ]
161
+ },
162
+ "count": {
163
+ "gain": 9.166666666666666,
164
+ "ci95": [
165
+ 4.323476942288824,
166
+ 14.361775596567533
167
+ ]
168
+ },
169
+ "behavior": {
170
+ "gain": 9.25,
171
+ "ci95": [
172
+ 5.482616495413512,
173
+ 13.075335023803161
174
+ ]
175
+ },
176
+ "runs": 3
177
+ },
178
+ "Middle8": {
179
+ "color": {
180
+ "gain": 12.166666666666666,
181
+ "ci95": [
182
+ 6.7306957186544345,
183
+ 17.894823265059202
184
+ ]
185
+ },
186
+ "count": {
187
+ "gain": 8.333333333333334,
188
+ "ci95": [
189
+ 3.2783364317503243,
190
+ 13.756986362620168
191
+ ]
192
+ },
193
+ "behavior": {
194
+ "gain": 10.25,
195
+ "ci95": [
196
+ 6.360450609418865,
197
+ 14.26668681251842
198
+ ]
199
+ },
200
+ "runs": 3
201
+ },
202
+ "Random8": {
203
+ "color": {
204
+ "gain": 9.95,
205
+ "ci95": [
206
+ 3.951123234916559,
207
+ 16.048490427098677
208
+ ]
209
+ },
210
+ "count": {
211
+ "gain": 8.7,
212
+ "ci95": [
213
+ 3.9285246821241584,
214
+ 14.077718205091049
215
+ ]
216
+ },
217
+ "behavior": {
218
+ "gain": 9.325,
219
+ "ci95": [
220
+ 5.38261732847098,
221
+ 13.281660126767152
222
+ ]
223
+ },
224
+ "runs": 10
225
+ },
226
+ "Shallow8": {
227
+ "color": {
228
+ "gain": 10.833333333333334,
229
+ "ci95": [
230
+ 5.102603605634744,
231
+ 16.838625063976018
232
+ ]
233
+ },
234
+ "count": {
235
+ "gain": 7.333333333333333,
236
+ "ci95": [
237
+ 1.9999231950844854,
238
+ 12.908993036466924
239
+ ]
240
+ },
241
+ "behavior": {
242
+ "gain": 9.083333333333334,
243
+ "ci95": [
244
+ 5.151902337103394,
245
+ 13.160007402628557
246
+ ]
247
+ },
248
+ "runs": 3
249
+ },
250
+ "Top8": {
251
+ "color": {
252
+ "gain": 10.5,
253
+ "ci95": [
254
+ 4.614942528735632,
255
+ 16.26428097167846
256
+ ]
257
+ },
258
+ "count": {
259
+ "gain": 7.833333333333333,
260
+ "ci95": [
261
+ 2.897635778070561,
262
+ 13.089161057092996
263
+ ]
264
+ },
265
+ "behavior": {
266
+ "gain": 9.166666666666666,
267
+ "ci95": [
268
+ 5.217074793047767,
269
+ 13.128772781465317
270
+ ]
271
+ },
272
+ "runs": 3
273
+ },
274
+ "TopDelta8": {
275
+ "color": {
276
+ "gain": 11.833333333333334,
277
+ "ci95": [
278
+ 6.292517006802721,
279
+ 17.51297738227661
280
+ ]
281
+ },
282
+ "count": {
283
+ "gain": 8.5,
284
+ "ci95": [
285
+ 3.7037037037037037,
286
+ 13.592741320591674
287
+ ]
288
+ },
289
+ "behavior": {
290
+ "gain": 10.166666666666666,
291
+ "ci95": [
292
+ 6.272588456853881,
293
+ 13.977588910057953
294
+ ]
295
+ },
296
+ "runs": 3
297
+ }
298
+ },
299
+ "contrasts": {
300
+ "Top8-Random8": {
301
+ "color": {
302
+ "difference": 0.5500000000000007,
303
+ "ci95": [
304
+ -2.181528835236759,
305
+ 3.200743853547795
306
+ ]
307
+ },
308
+ "count": {
309
+ "difference": -0.8666666666666663,
310
+ "ci95": [
311
+ -3.832101237864477,
312
+ 1.9146118721461194
313
+ ]
314
+ },
315
+ "behavior": {
316
+ "difference": -0.15833333333333321,
317
+ "ci95": [
318
+ -2.281880436956537,
319
+ 1.9427180643443083
320
+ ]
321
+ }
322
+ },
323
+ "TopDelta8-Random8": {
324
+ "color": {
325
+ "difference": 1.8833333333333346,
326
+ "ci95": [
327
+ -0.5586611938528271,
328
+ 4.410613529614792
329
+ ]
330
+ },
331
+ "count": {
332
+ "difference": -0.1999999999999993,
333
+ "ci95": [
334
+ -3.093110380116958,
335
+ 2.729087602783051
336
+ ]
337
+ },
338
+ "behavior": {
339
+ "difference": 0.8416666666666668,
340
+ "ci95": [
341
+ -0.9637530638165153,
342
+ 2.8065518555178453
343
+ ]
344
+ }
345
+ },
346
+ "Shallow8-Random8": {
347
+ "color": {
348
+ "difference": 0.8833333333333346,
349
+ "ci95": [
350
+ -2.079517704517706,
351
+ 4.029173252050881
352
+ ]
353
+ },
354
+ "count": {
355
+ "difference": -1.3666666666666663,
356
+ "ci95": [
357
+ -4.640565650343903,
358
+ 1.6243967392663345
359
+ ]
360
+ },
361
+ "behavior": {
362
+ "difference": -0.24166666666666536,
363
+ "ci95": [
364
+ -2.327714592834495,
365
+ 1.9401803985788748
366
+ ]
367
+ }
368
+ },
369
+ "Middle8-Random8": {
370
+ "color": {
371
+ "difference": 2.216666666666667,
372
+ "ci95": [
373
+ 0.055489417989417845,
374
+ 4.628924648223668
375
+ ]
376
+ },
377
+ "count": {
378
+ "difference": -0.36666666666666536,
379
+ "ci95": [
380
+ -3.621885738831616,
381
+ 2.7726371951219524
382
+ ]
383
+ },
384
+ "behavior": {
385
+ "difference": 0.9250000000000007,
386
+ "ci95": [
387
+ -1.18797866186717,
388
+ 3.0659404566424278
389
+ ]
390
+ }
391
+ },
392
+ "Deep8-Random8": {
393
+ "color": {
394
+ "difference": 0.21666666666666679,
395
+ "ci95": [
396
+ -2.68340210355987,
397
+ 3.155196264111707
398
+ ]
399
+ },
400
+ "count": {
401
+ "difference": -3.3666666666666663,
402
+ "ci95": [
403
+ -7.067083333333334,
404
+ 0.10530537573881361
405
+ ]
406
+ },
407
+ "behavior": {
408
+ "difference": -1.5749999999999993,
409
+ "ci95": [
410
+ -3.9747666664836414,
411
+ 0.7099735526578647
412
+ ]
413
+ }
414
+ },
415
+ "Top8-Shallow8": {
416
+ "color": {
417
+ "difference": -0.3333333333333339,
418
+ "ci95": [
419
+ -3.166666666666666,
420
+ 2.475247524752475
421
+ ]
422
+ },
423
+ "count": {
424
+ "difference": 0.5,
425
+ "ci95": [
426
+ -1.9419843713000227,
427
+ 3.4050179211469533
428
+ ]
429
+ },
430
+ "behavior": {
431
+ "difference": 0.08333333333333215,
432
+ "ci95": [
433
+ -1.8518795120736862,
434
+ 2.1669680102528024
435
+ ]
436
+ }
437
+ },
438
+ "Deep8-Full": {
439
+ "color": {
440
+ "difference": 0.8333333333333321,
441
+ "ci95": [
442
+ -2.3809523809523796,
443
+ 4.051799085733517
444
+ ]
445
+ },
446
+ "count": {
447
+ "difference": -3.833333333333333,
448
+ "ci95": [
449
+ -7.741522152807696,
450
+ -0.17262390082673582
451
+ ]
452
+ },
453
+ "behavior": {
454
+ "difference": -1.5,
455
+ "ci95": [
456
+ -3.862156197590915,
457
+ 0.8220405373870728
458
+ ]
459
+ }
460
+ },
461
+ "Middle8-Full": {
462
+ "color": {
463
+ "difference": 2.833333333333332,
464
+ "ci95": [
465
+ 0.341793570219965,
466
+ 5.666666666666666
467
+ ]
468
+ },
469
+ "count": {
470
+ "difference": -0.8333333333333321,
471
+ "ci95": [
472
+ -4.477851128970534,
473
+ 2.4062436253510975
474
+ ]
475
+ },
476
+ "behavior": {
477
+ "difference": 1.0,
478
+ "ci95": [
479
+ -1.2043076726577198,
480
+ 3.107975547150466
481
+ ]
482
+ }
483
+ },
484
+ "Random8-Full": {
485
+ "color": {
486
+ "difference": 0.6166666666666654,
487
+ "ci95": [
488
+ -1.3499999999999996,
489
+ 3.0694619185675487
490
+ ]
491
+ },
492
+ "count": {
493
+ "difference": -0.4666666666666668,
494
+ "ci95": [
495
+ -3.333758503401361,
496
+ 1.96538709875784
497
+ ]
498
+ },
499
+ "behavior": {
500
+ "difference": 0.07499999999999929,
501
+ "ci95": [
502
+ -1.629657520345539,
503
+ 1.512899304749076
504
+ ]
505
+ }
506
+ },
507
+ "Shallow8-Full": {
508
+ "color": {
509
+ "difference": 1.5,
510
+ "ci95": [
511
+ -1.629649758454108,
512
+ 4.975124378109451
513
+ ]
514
+ },
515
+ "count": {
516
+ "difference": -1.833333333333333,
517
+ "ci95": [
518
+ -5.218952194758646,
519
+ 1.258161126085656
520
+ ]
521
+ },
522
+ "behavior": {
523
+ "difference": -0.16666666666666607,
524
+ "ci95": [
525
+ -2.3435467387311233,
526
+ 1.995625026660409
527
+ ]
528
+ }
529
+ },
530
+ "Top8-Full": {
531
+ "color": {
532
+ "difference": 1.166666666666666,
533
+ "ci95": [
534
+ -1.7242123756774457,
535
+ 4.355108877721944
536
+ ]
537
+ },
538
+ "count": {
539
+ "difference": -1.333333333333333,
540
+ "ci95": [
541
+ -4.431624851391792,
542
+ 1.738383309029046
543
+ ]
544
+ },
545
+ "behavior": {
546
+ "difference": -0.08333333333333393,
547
+ "ci95": [
548
+ -2.152799599172603,
549
+ 2.0720516281469554
550
+ ]
551
+ }
552
+ },
553
+ "TopDelta8-Full": {
554
+ "color": {
555
+ "difference": 2.5,
556
+ "ci95": [
557
+ 1.7763568394002505e-15,
558
+ 5.034870426829269
559
+ ]
560
+ },
561
+ "count": {
562
+ "difference": -0.6666666666666661,
563
+ "ci95": [
564
+ -3.780161604903872,
565
+ 2.3150839793281666
566
+ ]
567
+ },
568
+ "behavior": {
569
+ "difference": 0.9166666666666661,
570
+ "ci95": [
571
+ -0.898544170848408,
572
+ 2.7615857987030425
573
+ ]
574
+ }
575
+ }
576
+ }
577
+ },
578
+ "fixed_internvl_replacement": {
579
+ "behavior": "replacement",
580
+ "arms": {
581
+ "Deep8": {
582
+ "color": {
583
+ "gain": 5.5,
584
+ "ci95": [
585
+ 2.1975709475709477,
586
+ 9.153808195258971
587
+ ]
588
+ },
589
+ "count": {
590
+ "gain": 0.6666666666666666,
591
+ "ci95": [
592
+ -4.283120510774142,
593
+ 5.699597314196522
594
+ ]
595
+ },
596
+ "behavior": {
597
+ "gain": 3.0833333333333335,
598
+ "ci95": [
599
+ 0.041972000238318305,
600
+ 6.1282139502268365
601
+ ]
602
+ },
603
+ "runs": 3
604
+ },
605
+ "Full": {
606
+ "color": {
607
+ "gain": 5.833333333333333,
608
+ "ci95": [
609
+ 2.4560113519091846,
610
+ 9.479161358270058
611
+ ]
612
+ },
613
+ "count": {
614
+ "gain": 0.6666666666666666,
615
+ "ci95": [
616
+ -4.715958061219957,
617
+ 6.250560035842296
618
+ ]
619
+ },
620
+ "behavior": {
621
+ "gain": 3.25,
622
+ "ci95": [
623
+ 0.03493230625583557,
624
+ 6.558021532610946
625
+ ]
626
+ },
627
+ "runs": 3
628
+ },
629
+ "Middle8": {
630
+ "color": {
631
+ "gain": 6.166666666666667,
632
+ "ci95": [
633
+ 2.640046899426785,
634
+ 10.047846889952153
635
+ ]
636
+ },
637
+ "count": {
638
+ "gain": 0.0,
639
+ "ci95": [
640
+ -5.5555555555555545,
641
+ 5.641093474426809
642
+ ]
643
+ },
644
+ "behavior": {
645
+ "gain": 3.0833333333333335,
646
+ "ci95": [
647
+ -0.1524001016193441,
648
+ 6.519581833336135
649
+ ]
650
+ },
651
+ "runs": 3
652
+ },
653
+ "Random8": {
654
+ "color": {
655
+ "gain": 5.05,
656
+ "ci95": [
657
+ 1.8817071223251354,
658
+ 8.398106796116503
659
+ ]
660
+ },
661
+ "count": {
662
+ "gain": 0.8,
663
+ "ci95": [
664
+ -4.0308084772370485,
665
+ 5.868597155069721
666
+ ]
667
+ },
668
+ "behavior": {
669
+ "gain": 2.925,
670
+ "ci95": [
671
+ -0.028902785810680506,
672
+ 5.880289649718909
673
+ ]
674
+ },
675
+ "runs": 10
676
+ },
677
+ "Shallow8": {
678
+ "color": {
679
+ "gain": 4.5,
680
+ "ci95": [
681
+ 1.0415317357512954,
682
+ 8.243819975612459
683
+ ]
684
+ },
685
+ "count": {
686
+ "gain": 0.3333333333333333,
687
+ "ci95": [
688
+ -5.263854079643553,
689
+ 6.122896527031868
690
+ ]
691
+ },
692
+ "behavior": {
693
+ "gain": 2.4166666666666665,
694
+ "ci95": [
695
+ -0.9788146698554657,
696
+ 5.777629470307291
697
+ ]
698
+ },
699
+ "runs": 3
700
+ },
701
+ "Top8": {
702
+ "color": {
703
+ "gain": 4.833333333333333,
704
+ "ci95": [
705
+ 1.4771185835653606,
706
+ 8.421908429610614
707
+ ]
708
+ },
709
+ "count": {
710
+ "gain": -0.16666666666666666,
711
+ "ci95": [
712
+ -5.729916557591623,
713
+ 5.492762445887447
714
+ ]
715
+ },
716
+ "behavior": {
717
+ "gain": 2.3333333333333335,
718
+ "ci95": [
719
+ -0.916828529657477,
720
+ 5.568640197353072
721
+ ]
722
+ },
723
+ "runs": 3
724
+ },
725
+ "TopDelta8": {
726
+ "color": {
727
+ "gain": 6.0,
728
+ "ci95": [
729
+ 2.512562814070352,
730
+ 9.667063492063493
731
+ ]
732
+ },
733
+ "count": {
734
+ "gain": 0.3333333333333333,
735
+ "ci95": [
736
+ -4.957264957264957,
737
+ 5.582943662324023
738
+ ]
739
+ },
740
+ "behavior": {
741
+ "gain": 3.1666666666666665,
742
+ "ci95": [
743
+ -0.07104606737497907,
744
+ 6.437948987509699
745
+ ]
746
+ },
747
+ "runs": 3
748
+ }
749
+ },
750
+ "contrasts": {
751
+ "Top8-Random8": {
752
+ "color": {
753
+ "difference": -0.21666666666666679,
754
+ "ci95": [
755
+ -1.7863636363636373,
756
+ 1.2592563621193162
757
+ ]
758
+ },
759
+ "count": {
760
+ "difference": -0.9666666666666667,
761
+ "ci95": [
762
+ -3.583524557165861,
763
+ 1.7464901988953845
764
+ ]
765
+ },
766
+ "behavior": {
767
+ "difference": -0.5916666666666663,
768
+ "ci95": [
769
+ -2.124358758652281,
770
+ 0.9655486008411758
771
+ ]
772
+ }
773
+ },
774
+ "TopDelta8-Random8": {
775
+ "color": {
776
+ "difference": 0.9500000000000002,
777
+ "ci95": [
778
+ -2.7755575615619447e-18,
779
+ 2.349755536381937
780
+ ]
781
+ },
782
+ "count": {
783
+ "difference": -0.46666666666666673,
784
+ "ci95": [
785
+ -2.4317695104337864,
786
+ 1.4085655235762413
787
+ ]
788
+ },
789
+ "behavior": {
790
+ "difference": 0.2416666666666667,
791
+ "ci95": [
792
+ -0.9317343049122525,
793
+ 1.551202012472187
794
+ ]
795
+ }
796
+ },
797
+ "Shallow8-Random8": {
798
+ "color": {
799
+ "difference": -0.5499999999999998,
800
+ "ci95": [
801
+ -2.4752475247524752,
802
+ 1.1836242138364783
803
+ ]
804
+ },
805
+ "count": {
806
+ "difference": -0.46666666666666673,
807
+ "ci95": [
808
+ -3.414695428361319,
809
+ 2.7127896613190745
810
+ ]
811
+ },
812
+ "behavior": {
813
+ "difference": -0.5083333333333333,
814
+ "ci95": [
815
+ -2.2997894154415555,
816
+ 1.3372060167433903
817
+ ]
818
+ }
819
+ },
820
+ "Middle8-Random8": {
821
+ "color": {
822
+ "difference": 1.1166666666666671,
823
+ "ci95": [
824
+ -0.338409475465312,
825
+ 2.947043066144653
826
+ ]
827
+ },
828
+ "count": {
829
+ "difference": -0.8,
830
+ "ci95": [
831
+ -3.4155755768732177,
832
+ 2.0574264394753117
833
+ ]
834
+ },
835
+ "behavior": {
836
+ "difference": 0.15833333333333366,
837
+ "ci95": [
838
+ -1.300025255127855,
839
+ 1.886065176539661
840
+ ]
841
+ }
842
+ },
843
+ "Deep8-Random8": {
844
+ "color": {
845
+ "difference": 0.4500000000000002,
846
+ "ci95": [
847
+ -0.6767266360782422,
848
+ 1.7058284667353703
849
+ ]
850
+ },
851
+ "count": {
852
+ "difference": -0.13333333333333341,
853
+ "ci95": [
854
+ -2.227644288819126,
855
+ 1.8501392804403283
856
+ ]
857
+ },
858
+ "behavior": {
859
+ "difference": 0.15833333333333366,
860
+ "ci95": [
861
+ -0.9475457822480527,
862
+ 1.2506931292087544
863
+ ]
864
+ }
865
+ },
866
+ "Top8-Shallow8": {
867
+ "color": {
868
+ "difference": 0.33333333333333304,
869
+ "ci95": [
870
+ -0.6472491909385107,
871
+ 1.4219065480565345
872
+ ]
873
+ },
874
+ "count": {
875
+ "difference": -0.5,
876
+ "ci95": [
877
+ -3.0895147095800852,
878
+ 1.8425460636515916
879
+ ]
880
+ },
881
+ "behavior": {
882
+ "difference": -0.08333333333333304,
883
+ "ci95": [
884
+ -1.4753264033196176,
885
+ 1.250011198196622
886
+ ]
887
+ }
888
+ },
889
+ "Deep8-Full": {
890
+ "color": {
891
+ "difference": -0.33333333333333304,
892
+ "ci95": [
893
+ -1.8518518518518512,
894
+ 0.9569377990430628
895
+ ]
896
+ },
897
+ "count": {
898
+ "difference": 0.0,
899
+ "ci95": [
900
+ -2.8709550757373354,
901
+ 2.4179620034542313
902
+ ]
903
+ },
904
+ "behavior": {
905
+ "difference": -0.16666666666666652,
906
+ "ci95": [
907
+ -1.8273203478127629,
908
+ 1.1345159142640158
909
+ ]
910
+ }
911
+ },
912
+ "Middle8-Full": {
913
+ "color": {
914
+ "difference": 0.3333333333333339,
915
+ "ci95": [
916
+ -1.2404549735564718,
917
+ 2.1166835528537673
918
+ ]
919
+ },
920
+ "count": {
921
+ "difference": -0.6666666666666666,
922
+ "ci95": [
923
+ -3.608450042144849,
924
+ 2.4563131967850667
925
+ ]
926
+ },
927
+ "behavior": {
928
+ "difference": -0.16666666666666652,
929
+ "ci95": [
930
+ -1.7780610883128138,
931
+ 1.6718349602314826
932
+ ]
933
+ }
934
+ },
935
+ "Random8-Full": {
936
+ "color": {
937
+ "difference": -0.7833333333333332,
938
+ "ci95": [
939
+ -2.1539682539682525,
940
+ 0.09393024815559997
941
+ ]
942
+ },
943
+ "count": {
944
+ "difference": 0.13333333333333341,
945
+ "ci95": [
946
+ -2.425782770584751,
947
+ 2.3811846689895484
948
+ ]
949
+ },
950
+ "behavior": {
951
+ "difference": -0.3250000000000002,
952
+ "ci95": [
953
+ -1.9248199954998866,
954
+ 0.8313538662677991
955
+ ]
956
+ }
957
+ },
958
+ "Shallow8-Full": {
959
+ "color": {
960
+ "difference": -1.333333333333333,
961
+ "ci95": [
962
+ -3.472304502369669,
963
+ 0.49757462686567183
964
+ ]
965
+ },
966
+ "count": {
967
+ "difference": -0.3333333333333333,
968
+ "ci95": [
969
+ -3.3874103552626122,
970
+ 2.966841186736475
971
+ ]
972
+ },
973
+ "behavior": {
974
+ "difference": -0.8333333333333335,
975
+ "ci95": [
976
+ -2.727315779882576,
977
+ 1.011495180903665
978
+ ]
979
+ }
980
+ },
981
+ "Top8-Full": {
982
+ "color": {
983
+ "difference": -1.0,
984
+ "ci95": [
985
+ -2.837047494201875,
986
+ 0.5025125628140703
987
+ ]
988
+ },
989
+ "count": {
990
+ "difference": -0.8333333333333333,
991
+ "ci95": [
992
+ -3.426791277258567,
993
+ 1.9906648078289892
994
+ ]
995
+ },
996
+ "behavior": {
997
+ "difference": -0.9166666666666665,
998
+ "ci95": [
999
+ -2.5295529286117837,
1000
+ 0.7356618762855107
1001
+ ]
1002
+ }
1003
+ },
1004
+ "TopDelta8-Full": {
1005
+ "color": {
1006
+ "difference": 0.16666666666666696,
1007
+ "ci95": [
1008
+ -1.0002525252525252,
1009
+ 1.4563327851168058
1010
+ ]
1011
+ },
1012
+ "count": {
1013
+ "difference": -0.3333333333333333,
1014
+ "ci95": [
1015
+ -2.724358974358974,
1016
+ 1.8730125079499695
1017
+ ]
1018
+ },
1019
+ "behavior": {
1020
+ "difference": -0.08333333333333348,
1021
+ "ci95": [
1022
+ -1.5632915846466247,
1023
+ 1.261240453313624
1024
+ ]
1025
+ }
1026
+ }
1027
+ }
1028
+ },
1029
+ "fixed_internvl_context": {
1030
+ "behavior": "context",
1031
+ "arms": {
1032
+ "Deep8": {
1033
+ "color": {
1034
+ "gain": 0.6666666666666666,
1035
+ "ci95": [
1036
+ -2.8851495726495724,
1037
+ 4.30134274862248
1038
+ ]
1039
+ },
1040
+ "count": {
1041
+ "gain": 6.0,
1042
+ "ci95": [
1043
+ -0.3089387464387463,
1044
+ 12.440634414318627
1045
+ ]
1046
+ },
1047
+ "behavior": {
1048
+ "gain": 3.3333333333333335,
1049
+ "ci95": [
1050
+ -0.14629087479152567,
1051
+ 6.896319368090022
1052
+ ]
1053
+ },
1054
+ "runs": 3
1055
+ },
1056
+ "Full": {
1057
+ "color": {
1058
+ "gain": -1.3333333333333333,
1059
+ "ci95": [
1060
+ -5.235845713627626,
1061
+ 2.525252525252525
1062
+ ]
1063
+ },
1064
+ "count": {
1065
+ "gain": 6.166666666666667,
1066
+ "ci95": [
1067
+ 0.1593250223139166,
1068
+ 12.439365671641797
1069
+ ]
1070
+ },
1071
+ "behavior": {
1072
+ "gain": 2.4166666666666665,
1073
+ "ci95": [
1074
+ -1.0492949917524397,
1075
+ 6.024416320892247
1076
+ ]
1077
+ },
1078
+ "runs": 3
1079
+ },
1080
+ "Middle8": {
1081
+ "color": {
1082
+ "gain": -1.1666666666666667,
1083
+ "ci95": [
1084
+ -5.090369002006933,
1085
+ 2.551677887649907
1086
+ ]
1087
+ },
1088
+ "count": {
1089
+ "gain": 5.833333333333333,
1090
+ "ci95": [
1091
+ -0.15515003152702056,
1092
+ 11.800633309566754
1093
+ ]
1094
+ },
1095
+ "behavior": {
1096
+ "gain": 2.3333333333333335,
1097
+ "ci95": [
1098
+ -1.1875944213336809,
1099
+ 5.952674244050197
1100
+ ]
1101
+ },
1102
+ "runs": 3
1103
+ },
1104
+ "Random8": {
1105
+ "color": {
1106
+ "gain": -1.35,
1107
+ "ci95": [
1108
+ -5.240502450980392,
1109
+ 2.376411122420653
1110
+ ]
1111
+ },
1112
+ "count": {
1113
+ "gain": 5.5,
1114
+ "ci95": [
1115
+ -0.5127975549332181,
1116
+ 11.65011934386268
1117
+ ]
1118
+ },
1119
+ "behavior": {
1120
+ "gain": 2.075,
1121
+ "ci95": [
1122
+ -1.4575639045988473,
1123
+ 5.653853078777224
1124
+ ]
1125
+ },
1126
+ "runs": 10
1127
+ },
1128
+ "Shallow8": {
1129
+ "color": {
1130
+ "gain": -1.6666666666666667,
1131
+ "ci95": [
1132
+ -5.650520030816641,
1133
+ 2.041077969649399
1134
+ ]
1135
+ },
1136
+ "count": {
1137
+ "gain": 8.0,
1138
+ "ci95": [
1139
+ 1.1709938581797883,
1140
+ 15.13607496138552
1141
+ ]
1142
+ },
1143
+ "behavior": {
1144
+ "gain": 3.1666666666666665,
1145
+ "ci95": [
1146
+ -0.7516368938263859,
1147
+ 7.30842545834873
1148
+ ]
1149
+ },
1150
+ "runs": 3
1151
+ },
1152
+ "Top8": {
1153
+ "color": {
1154
+ "gain": -1.1666666666666667,
1155
+ "ci95": [
1156
+ -5.0,
1157
+ 2.6604251737230493
1158
+ ]
1159
+ },
1160
+ "count": {
1161
+ "gain": 7.333333333333333,
1162
+ "ci95": [
1163
+ 1.2930001198609617,
1164
+ 13.70608141733851
1165
+ ]
1166
+ },
1167
+ "behavior": {
1168
+ "gain": 3.0833333333333335,
1169
+ "ci95": [
1170
+ -0.38387615699588057,
1171
+ 6.626760928184765
1172
+ ]
1173
+ },
1174
+ "runs": 3
1175
+ },
1176
+ "TopDelta8": {
1177
+ "color": {
1178
+ "gain": -1.0,
1179
+ "ci95": [
1180
+ -5.009047829822689,
1181
+ 2.7178325901152003
1182
+ ]
1183
+ },
1184
+ "count": {
1185
+ "gain": 6.166666666666667,
1186
+ "ci95": [
1187
+ 0.0,
1188
+ 12.661406765676572
1189
+ ]
1190
+ },
1191
+ "behavior": {
1192
+ "gain": 2.5833333333333335,
1193
+ "ci95": [
1194
+ -1.0620537828110266,
1195
+ 6.287815742174594
1196
+ ]
1197
+ },
1198
+ "runs": 3
1199
+ }
1200
+ },
1201
+ "contrasts": {
1202
+ "Top8-Random8": {
1203
+ "color": {
1204
+ "difference": 0.18333333333333335,
1205
+ "ci95": [
1206
+ -2.160848960248397,
1207
+ 2.364025682736864
1208
+ ]
1209
+ },
1210
+ "count": {
1211
+ "difference": 1.833333333333333,
1212
+ "ci95": [
1213
+ -1.2166732804232814,
1214
+ 5.50582260267291
1215
+ ]
1216
+ },
1217
+ "behavior": {
1218
+ "difference": 1.0083333333333333,
1219
+ "ci95": [
1220
+ -0.8398903832366266,
1221
+ 2.9564089965896843
1222
+ ]
1223
+ }
1224
+ },
1225
+ "TopDelta8-Random8": {
1226
+ "color": {
1227
+ "difference": 0.3500000000000001,
1228
+ "ci95": [
1229
+ -1.4716913468049464,
1230
+ 2.1976762207848584
1231
+ ]
1232
+ },
1233
+ "count": {
1234
+ "difference": 0.666666666666667,
1235
+ "ci95": [
1236
+ -2.492174857724686,
1237
+ 4.010644745814078
1238
+ ]
1239
+ },
1240
+ "behavior": {
1241
+ "difference": 0.5083333333333333,
1242
+ "ci95": [
1243
+ -1.4081065709933367,
1244
+ 2.507500587254224
1245
+ ]
1246
+ }
1247
+ },
1248
+ "Shallow8-Random8": {
1249
+ "color": {
1250
+ "difference": -0.31666666666666665,
1251
+ "ci95": [
1252
+ -3.317155993278566,
1253
+ 2.921225885475147
1254
+ ]
1255
+ },
1256
+ "count": {
1257
+ "difference": 2.5,
1258
+ "ci95": [
1259
+ -2.0037219970730615,
1260
+ 7.081490686115112
1261
+ ]
1262
+ },
1263
+ "behavior": {
1264
+ "difference": 1.0916666666666663,
1265
+ "ci95": [
1266
+ -1.890742755680293,
1267
+ 4.073317216367707
1268
+ ]
1269
+ }
1270
+ },
1271
+ "Middle8-Random8": {
1272
+ "color": {
1273
+ "difference": 0.18333333333333335,
1274
+ "ci95": [
1275
+ -2.1770606300772735,
1276
+ 2.1987757514353268
1277
+ ]
1278
+ },
1279
+ "count": {
1280
+ "difference": 0.33333333333333304,
1281
+ "ci95": [
1282
+ -2.908586511527689,
1283
+ 3.6753663003663
1284
+ ]
1285
+ },
1286
+ "behavior": {
1287
+ "difference": 0.2583333333333333,
1288
+ "ci95": [
1289
+ -1.876531954887218,
1290
+ 2.247400161900185
1291
+ ]
1292
+ }
1293
+ },
1294
+ "Deep8-Random8": {
1295
+ "color": {
1296
+ "difference": 2.0166666666666666,
1297
+ "ci95": [
1298
+ -0.862992179997256,
1299
+ 5.200811991162236
1300
+ ]
1301
+ },
1302
+ "count": {
1303
+ "difference": 0.5,
1304
+ "ci95": [
1305
+ -3.760459056186194,
1306
+ 4.877094393164028
1307
+ ]
1308
+ },
1309
+ "behavior": {
1310
+ "difference": 1.2583333333333333,
1311
+ "ci95": [
1312
+ -1.2798553307024108,
1313
+ 3.905312345388439
1314
+ ]
1315
+ }
1316
+ },
1317
+ "Top8-Shallow8": {
1318
+ "color": {
1319
+ "difference": 0.5,
1320
+ "ci95": [
1321
+ -2.623659844054581,
1322
+ 3.4017966160823323
1323
+ ]
1324
+ },
1325
+ "count": {
1326
+ "difference": -0.666666666666667,
1327
+ "ci95": [
1328
+ -5.664223219658161,
1329
+ 4.592425431711148
1330
+ ]
1331
+ },
1332
+ "behavior": {
1333
+ "difference": -0.08333333333333304,
1334
+ "ci95": [
1335
+ -3.2126736137582186,
1336
+ 2.9376729918440287
1337
+ ]
1338
+ }
1339
+ },
1340
+ "Deep8-Full": {
1341
+ "color": {
1342
+ "difference": 2.0,
1343
+ "ci95": [
1344
+ -0.9760951340856691,
1345
+ 5.263381668456858
1346
+ ]
1347
+ },
1348
+ "count": {
1349
+ "difference": -0.16666666666666696,
1350
+ "ci95": [
1351
+ -4.598006847640008,
1352
+ 4.146455223880598
1353
+ ]
1354
+ },
1355
+ "behavior": {
1356
+ "difference": 0.916666666666667,
1357
+ "ci95": [
1358
+ -1.5576810891528263,
1359
+ 3.418970352564103
1360
+ ]
1361
+ }
1362
+ },
1363
+ "Middle8-Full": {
1364
+ "color": {
1365
+ "difference": 0.16666666666666652,
1366
+ "ci95": [
1367
+ -2.427480464125029,
1368
+ 2.5516571362844678
1369
+ ]
1370
+ },
1371
+ "count": {
1372
+ "difference": -0.3333333333333339,
1373
+ "ci95": [
1374
+ -4.085338680926918,
1375
+ 3.2815836715247113
1376
+ ]
1377
+ },
1378
+ "behavior": {
1379
+ "difference": -0.08333333333333304,
1380
+ "ci95": [
1381
+ -2.33340673272612,
1382
+ 1.9279814647004245
1383
+ ]
1384
+ }
1385
+ },
1386
+ "Random8-Full": {
1387
+ "color": {
1388
+ "difference": -0.01666666666666683,
1389
+ "ci95": [
1390
+ -1.8790123917587334,
1391
+ 1.7173896323642523
1392
+ ]
1393
+ },
1394
+ "count": {
1395
+ "difference": -0.666666666666667,
1396
+ "ci95": [
1397
+ -3.688064942288972,
1398
+ 2.397506105540414
1399
+ ]
1400
+ },
1401
+ "behavior": {
1402
+ "difference": -0.34166666666666634,
1403
+ "ci95": [
1404
+ -1.9993687147141845,
1405
+ 1.3131350530382426
1406
+ ]
1407
+ }
1408
+ },
1409
+ "Shallow8-Full": {
1410
+ "color": {
1411
+ "difference": -0.3333333333333335,
1412
+ "ci95": [
1413
+ -3.465777547904044,
1414
+ 3.0748924031628486
1415
+ ]
1416
+ },
1417
+ "count": {
1418
+ "difference": 1.833333333333333,
1419
+ "ci95": [
1420
+ -3.151305970149254,
1421
+ 7.193224896411884
1422
+ ]
1423
+ },
1424
+ "behavior": {
1425
+ "difference": 0.75,
1426
+ "ci95": [
1427
+ -2.329369379284607,
1428
+ 3.9194275271084056
1429
+ ]
1430
+ }
1431
+ },
1432
+ "Top8-Full": {
1433
+ "color": {
1434
+ "difference": 0.16666666666666652,
1435
+ "ci95": [
1436
+ -2.39298493525382,
1437
+ 2.4756716325838206
1438
+ ]
1439
+ },
1440
+ "count": {
1441
+ "difference": 1.166666666666666,
1442
+ "ci95": [
1443
+ -2.469596462133776,
1444
+ 5.1158288905813665
1445
+ ]
1446
+ },
1447
+ "behavior": {
1448
+ "difference": 0.666666666666667,
1449
+ "ci95": [
1450
+ -1.4385378065395786,
1451
+ 2.642019828518
1452
+ ]
1453
+ }
1454
+ },
1455
+ "TopDelta8-Full": {
1456
+ "color": {
1457
+ "difference": 0.33333333333333326,
1458
+ "ci95": [
1459
+ -1.5708872967759713,
1460
+ 2.1835347654124955
1461
+ ]
1462
+ },
1463
+ "count": {
1464
+ "difference": 0.0,
1465
+ "ci95": [
1466
+ -3.61292656359232,
1467
+ 3.6088003130209954
1468
+ ]
1469
+ },
1470
+ "behavior": {
1471
+ "difference": 0.16666666666666696,
1472
+ "ci95": [
1473
+ -1.8426120250215976,
1474
+ 2.082723897389739
1475
+ ]
1476
+ }
1477
+ }
1478
+ }
1479
+ },
1480
+ "single_internvl_removal": {
1481
+ "behavior": "removal",
1482
+ "runs": [
1483
+ {
1484
+ "layer": 1,
1485
+ "seed": 13,
1486
+ "color": {
1487
+ "items": 200,
1488
+ "gain": -0.5,
1489
+ "ci95": [
1490
+ -4.371584699453552,
1491
+ 3.5916230366492043
1492
+ ]
1493
+ },
1494
+ "count": {
1495
+ "items": 200,
1496
+ "gain": 5.0,
1497
+ "ci95": [
1498
+ 0.0,
1499
+ 9.95260663507109
1500
+ ]
1501
+ },
1502
+ "behavior": {
1503
+ "gain": 2.25,
1504
+ "ci95": [
1505
+ -0.8067265712179895,
1506
+ 5.422825376773071
1507
+ ]
1508
+ }
1509
+ },
1510
+ {
1511
+ "layer": 2,
1512
+ "seed": 13,
1513
+ "color": {
1514
+ "items": 200,
1515
+ "gain": 1.0,
1516
+ "ci95": [
1517
+ -2.689149560117302,
1518
+ 5.025644276428009
1519
+ ]
1520
+ },
1521
+ "count": {
1522
+ "items": 200,
1523
+ "gain": 3.5000000000000004,
1524
+ "ci95": [
1525
+ -1.0418030104712042,
1526
+ 8.45087940578796
1527
+ ]
1528
+ },
1529
+ "behavior": {
1530
+ "gain": 2.25,
1531
+ "ci95": [
1532
+ -0.741377665288737,
1533
+ 5.354181079244707
1534
+ ]
1535
+ }
1536
+ },
1537
+ {
1538
+ "layer": 7,
1539
+ "seed": 13,
1540
+ "color": {
1541
+ "items": 200,
1542
+ "gain": 8.5,
1543
+ "ci95": [
1544
+ 2.2337160046657254,
1545
+ 14.438811783252449
1546
+ ]
1547
+ },
1548
+ "count": {
1549
+ "items": 200,
1550
+ "gain": 5.0,
1551
+ "ci95": [
1552
+ 0.0,
1553
+ 10.309278350515465
1554
+ ]
1555
+ },
1556
+ "behavior": {
1557
+ "gain": 6.75,
1558
+ "ci95": [
1559
+ 2.7626606399757736,
1560
+ 10.758627872651084
1561
+ ]
1562
+ }
1563
+ },
1564
+ {
1565
+ "layer": 9,
1566
+ "seed": 13,
1567
+ "color": {
1568
+ "items": 200,
1569
+ "gain": 5.0,
1570
+ "ci95": [
1571
+ -1.0050251256281406,
1572
+ 10.796186738519209
1573
+ ]
1574
+ },
1575
+ "count": {
1576
+ "items": 200,
1577
+ "gain": 6.0,
1578
+ "ci95": [
1579
+ 1.0415317357512954,
1580
+ 11.112403100775186
1581
+ ]
1582
+ },
1583
+ "behavior": {
1584
+ "gain": 5.5,
1585
+ "ci95": [
1586
+ 1.6650991548451843,
1587
+ 9.48074463899272
1588
+ ]
1589
+ }
1590
+ },
1591
+ {
1592
+ "layer": 11,
1593
+ "seed": 13,
1594
+ "color": {
1595
+ "items": 200,
1596
+ "gain": 9.0,
1597
+ "ci95": [
1598
+ 3.4995646766169153,
1599
+ 14.438811783252449
1600
+ ]
1601
+ },
1602
+ "count": {
1603
+ "items": 200,
1604
+ "gain": 5.5,
1605
+ "ci95": [
1606
+ 0.5460847575863272,
1607
+ 10.312209898861568
1608
+ ]
1609
+ },
1610
+ "behavior": {
1611
+ "gain": 7.25,
1612
+ "ci95": [
1613
+ 3.7335642023222713,
1614
+ 10.937554178224687
1615
+ ]
1616
+ }
1617
+ },
1618
+ {
1619
+ "layer": 16,
1620
+ "seed": 13,
1621
+ "color": {
1622
+ "items": 200,
1623
+ "gain": 7.000000000000001,
1624
+ "ci95": [
1625
+ 0.9998756218905472,
1626
+ 13.170731707317072
1627
+ ]
1628
+ },
1629
+ "count": {
1630
+ "items": 200,
1631
+ "gain": 7.5,
1632
+ "ci95": [
1633
+ 2.55069667460893,
1634
+ 12.821428571428568
1635
+ ]
1636
+ },
1637
+ "behavior": {
1638
+ "gain": 7.25,
1639
+ "ci95": [
1640
+ 3.473338104337174,
1641
+ 11.41884378260093
1642
+ ]
1643
+ }
1644
+ },
1645
+ {
1646
+ "layer": 20,
1647
+ "seed": 13,
1648
+ "color": {
1649
+ "items": 200,
1650
+ "gain": 8.5,
1651
+ "ci95": [
1652
+ 2.984339386809794,
1653
+ 14.045579018688946
1654
+ ]
1655
+ },
1656
+ "count": {
1657
+ "items": 200,
1658
+ "gain": 8.5,
1659
+ "ci95": [
1660
+ 3.979119182413058,
1661
+ 13.334224598930478
1662
+ ]
1663
+ },
1664
+ "behavior": {
1665
+ "gain": 8.5,
1666
+ "ci95": [
1667
+ 4.9918503381439425,
1668
+ 12.470606629143212
1669
+ ]
1670
+ }
1671
+ },
1672
+ {
1673
+ "layer": 25,
1674
+ "seed": 13,
1675
+ "color": {
1676
+ "items": 200,
1677
+ "gain": 7.000000000000001,
1678
+ "ci95": [
1679
+ 1.075125064688632,
1680
+ 13.298758865248221
1681
+ ]
1682
+ },
1683
+ "count": {
1684
+ "items": 200,
1685
+ "gain": 7.5,
1686
+ "ci95": [
1687
+ 1.5147727272727274,
1688
+ 13.021399456521737
1689
+ ]
1690
+ },
1691
+ "behavior": {
1692
+ "gain": 7.25,
1693
+ "ci95": [
1694
+ 3.1573756909695843,
1695
+ 11.548631821036997
1696
+ ]
1697
+ }
1698
+ },
1699
+ {
1700
+ "layer": 29,
1701
+ "seed": 13,
1702
+ "color": {
1703
+ "items": 200,
1704
+ "gain": 7.5,
1705
+ "ci95": [
1706
+ 2.0613296865137807,
1707
+ 13.208046321253866
1708
+ ]
1709
+ },
1710
+ "count": {
1711
+ "items": 200,
1712
+ "gain": 7.5,
1713
+ "ci95": [
1714
+ 1.8957345971563981,
1715
+ 12.82079646017699
1716
+ ]
1717
+ },
1718
+ "behavior": {
1719
+ "gain": 7.5,
1720
+ "ci95": [
1721
+ 3.544871871150046,
1722
+ 11.536953773471842
1723
+ ]
1724
+ }
1725
+ },
1726
+ {
1727
+ "layer": 34,
1728
+ "seed": 13,
1729
+ "color": {
1730
+ "items": 200,
1731
+ "gain": 9.0,
1732
+ "ci95": [
1733
+ 4.2105263157894735,
1734
+ 14.079263508875151
1735
+ ]
1736
+ },
1737
+ "count": {
1738
+ "items": 200,
1739
+ "gain": 3.5000000000000004,
1740
+ "ci95": [
1741
+ -1.5625,
1742
+ 8.629441624365482
1743
+ ]
1744
+ },
1745
+ "behavior": {
1746
+ "gain": 6.25,
1747
+ "ci95": [
1748
+ 2.6447498748085305,
1749
+ 9.89256471681922
1750
+ ]
1751
+ }
1752
+ }
1753
+ ],
1754
+ "layers": [
1755
+ 1,
1756
+ 2,
1757
+ 7,
1758
+ 9,
1759
+ 11,
1760
+ 16,
1761
+ 20,
1762
+ 25,
1763
+ 29,
1764
+ 34
1765
+ ],
1766
+ "layer_gain": [
1767
+ 2.25,
1768
+ 2.25,
1769
+ 6.75,
1770
+ 5.5,
1771
+ 7.25,
1772
+ 7.25,
1773
+ 8.5,
1774
+ 7.25,
1775
+ 7.5,
1776
+ 6.25
1777
+ ],
1778
+ "spearman_gamma_gain": -0.5662208585049305
1779
+ },
1780
+ "single_internvl_replacement": {
1781
+ "behavior": "replacement",
1782
+ "runs": [
1783
+ {
1784
+ "layer": 1,
1785
+ "seed": 13,
1786
+ "color": {
1787
+ "items": 200,
1788
+ "gain": 3.5000000000000004,
1789
+ "ci95": [
1790
+ 0.9659524304833229,
1791
+ 6.566489770804488
1792
+ ]
1793
+ },
1794
+ "count": {
1795
+ "items": 200,
1796
+ "gain": -2.0,
1797
+ "ci95": [
1798
+ -6.418376034267166,
1799
+ 2.0730814095434416
1800
+ ]
1801
+ },
1802
+ "behavior": {
1803
+ "gain": 0.7500000000000002,
1804
+ "ci95": [
1805
+ -1.904864949776407,
1806
+ 3.3343535310393158
1807
+ ]
1808
+ }
1809
+ },
1810
+ {
1811
+ "layer": 2,
1812
+ "seed": 13,
1813
+ "color": {
1814
+ "items": 200,
1815
+ "gain": 4.0,
1816
+ "ci95": [
1817
+ 1.0415317357512954,
1818
+ 7.281553398058253
1819
+ ]
1820
+ },
1821
+ "count": {
1822
+ "items": 200,
1823
+ "gain": -2.5,
1824
+ "ci95": [
1825
+ -6.372549019607843,
1826
+ 1.5003787878787858
1827
+ ]
1828
+ },
1829
+ "behavior": {
1830
+ "gain": 0.75,
1831
+ "ci95": [
1832
+ -1.7565732484240026,
1833
+ 3.340770753133655
1834
+ ]
1835
+ }
1836
+ },
1837
+ {
1838
+ "layer": 7,
1839
+ "seed": 13,
1840
+ "color": {
1841
+ "items": 200,
1842
+ "gain": 4.0,
1843
+ "ci95": [
1844
+ 1.0415317357512954,
1845
+ 7.281553398058253
1846
+ ]
1847
+ },
1848
+ "count": {
1849
+ "items": 200,
1850
+ "gain": -1.0,
1851
+ "ci95": [
1852
+ -5.612244897959184,
1853
+ 3.333732057416266
1854
+ ]
1855
+ },
1856
+ "behavior": {
1857
+ "gain": 1.5,
1858
+ "ci95": [
1859
+ -1.2689268895210943,
1860
+ 4.266941740931832
1861
+ ]
1862
+ }
1863
+ },
1864
+ {
1865
+ "layer": 9,
1866
+ "seed": 13,
1867
+ "color": {
1868
+ "items": 200,
1869
+ "gain": 4.5,
1870
+ "ci95": [
1871
+ 1.5075376884422111,
1872
+ 8.001005025125624
1873
+ ]
1874
+ },
1875
+ "count": {
1876
+ "items": 200,
1877
+ "gain": -1.0,
1878
+ "ci95": [
1879
+ -6.122448979591836,
1880
+ 3.865676917638736
1881
+ ]
1882
+ },
1883
+ "behavior": {
1884
+ "gain": 1.75,
1885
+ "ci95": [
1886
+ -1.115608798189774,
1887
+ 4.709702006909681
1888
+ ]
1889
+ }
1890
+ },
1891
+ {
1892
+ "layer": 11,
1893
+ "seed": 13,
1894
+ "color": {
1895
+ "items": 200,
1896
+ "gain": 3.5000000000000004,
1897
+ "ci95": [
1898
+ 0.5492277992277993,
1899
+ 6.63265306122449
1900
+ ]
1901
+ },
1902
+ "count": {
1903
+ "items": 200,
1904
+ "gain": -1.0,
1905
+ "ci95": [
1906
+ -6.350050658561297,
1907
+ 4.040404040404041
1908
+ ]
1909
+ },
1910
+ "behavior": {
1911
+ "gain": 1.2500000000000002,
1912
+ "ci95": [
1913
+ -1.6821572564404461,
1914
+ 4.2092315938910945
1915
+ ]
1916
+ }
1917
+ },
1918
+ {
1919
+ "layer": 16,
1920
+ "seed": 13,
1921
+ "color": {
1922
+ "items": 200,
1923
+ "gain": 6.0,
1924
+ "ci95": [
1925
+ 2.5122487437185934,
1926
+ 9.743589743589743
1927
+ ]
1928
+ },
1929
+ "count": {
1930
+ "items": 200,
1931
+ "gain": 0.5,
1932
+ "ci95": [
1933
+ -4.976374409360234,
1934
+ 5.7976765585461205
1935
+ ]
1936
+ },
1937
+ "behavior": {
1938
+ "gain": 3.25,
1939
+ "ci95": [
1940
+ 0.07943779537896004,
1941
+ 6.524006038701953
1942
+ ]
1943
+ }
1944
+ },
1945
+ {
1946
+ "layer": 20,
1947
+ "seed": 13,
1948
+ "color": {
1949
+ "items": 200,
1950
+ "gain": 5.5,
1951
+ "ci95": [
1952
+ 2.116123642439432,
1953
+ 9.090909090909092
1954
+ ]
1955
+ },
1956
+ "count": {
1957
+ "items": 200,
1958
+ "gain": 1.0,
1959
+ "ci95": [
1960
+ -4.478171641791044,
1961
+ 6.436444017536078
1962
+ ]
1963
+ },
1964
+ "behavior": {
1965
+ "gain": 3.25,
1966
+ "ci95": [
1967
+ -0.0009716893603148982,
1968
+ 6.62662292928895
1969
+ ]
1970
+ }
1971
+ },
1972
+ {
1973
+ "layer": 25,
1974
+ "seed": 13,
1975
+ "color": {
1976
+ "items": 200,
1977
+ "gain": 4.5,
1978
+ "ci95": [
1979
+ 1.4851485148514851,
1980
+ 7.870655446086479
1981
+ ]
1982
+ },
1983
+ "count": {
1984
+ "items": 200,
1985
+ "gain": 0.5,
1986
+ "ci95": [
1987
+ -4.976374409360234,
1988
+ 5.7976765585461205
1989
+ ]
1990
+ },
1991
+ "behavior": {
1992
+ "gain": 2.5,
1993
+ "ci95": [
1994
+ -0.6082272139162945,
1995
+ 5.660514010539885
1996
+ ]
1997
+ }
1998
+ },
1999
+ {
2000
+ "layer": 29,
2001
+ "seed": 13,
2002
+ "color": {
2003
+ "items": 200,
2004
+ "gain": 3.5000000000000004,
2005
+ "ci95": [
2006
+ 0.8766440131032127,
2007
+ 6.567331478045754
2008
+ ]
2009
+ },
2010
+ "count": {
2011
+ "items": 200,
2012
+ "gain": -1.5,
2013
+ "ci95": [
2014
+ -6.666666666666667,
2015
+ 3.3828857630008455
2016
+ ]
2017
+ },
2018
+ "behavior": {
2019
+ "gain": 1.0000000000000002,
2020
+ "ci95": [
2021
+ -1.972781046881313,
2022
+ 3.8965921868572906
2023
+ ]
2024
+ }
2025
+ },
2026
+ {
2027
+ "layer": 34,
2028
+ "seed": 13,
2029
+ "color": {
2030
+ "items": 200,
2031
+ "gain": 3.5000000000000004,
2032
+ "ci95": [
2033
+ 0.8766440131032127,
2034
+ 6.567331478045754
2035
+ ]
2036
+ },
2037
+ "count": {
2038
+ "items": 200,
2039
+ "gain": -0.5,
2040
+ "ci95": [
2041
+ -4.787234042553192,
2042
+ 3.684271161775406
2043
+ ]
2044
+ },
2045
+ "behavior": {
2046
+ "gain": 1.5000000000000002,
2047
+ "ci95": [
2048
+ -1.0660448389882267,
2049
+ 4.095679921546328
2050
+ ]
2051
+ }
2052
+ }
2053
+ ],
2054
+ "layers": [
2055
+ 1,
2056
+ 2,
2057
+ 7,
2058
+ 9,
2059
+ 11,
2060
+ 16,
2061
+ 20,
2062
+ 25,
2063
+ 29,
2064
+ 34
2065
+ ],
2066
+ "layer_gain": [
2067
+ 0.7500000000000002,
2068
+ 0.75,
2069
+ 1.5,
2070
+ 1.75,
2071
+ 1.2500000000000002,
2072
+ 3.25,
2073
+ 3.25,
2074
+ 2.5,
2075
+ 1.0000000000000002,
2076
+ 1.5000000000000002
2077
+ ],
2078
+ "spearman_gamma_gain": -0.4012176588288426
2079
+ },
2080
+ "single_internvl_context": {
2081
+ "behavior": "context",
2082
+ "runs": [
2083
+ {
2084
+ "layer": 1,
2085
+ "seed": 13,
2086
+ "color": {
2087
+ "items": 200,
2088
+ "gain": -1.0,
2089
+ "ci95": [
2090
+ -4.285714285714286,
2091
+ 2.0103040454799235
2092
+ ]
2093
+ },
2094
+ "count": {
2095
+ "items": 200,
2096
+ "gain": 6.5,
2097
+ "ci95": [
2098
+ 0.0,
2099
+ 13.461787564766839
2100
+ ]
2101
+ },
2102
+ "behavior": {
2103
+ "gain": 2.75,
2104
+ "ci95": [
2105
+ -0.9800188879127214,
2106
+ 6.555780578057806
2107
+ ]
2108
+ }
2109
+ },
2110
+ {
2111
+ "layer": 2,
2112
+ "seed": 13,
2113
+ "color": {
2114
+ "items": 200,
2115
+ "gain": -0.5,
2116
+ "ci95": [
2117
+ -3.0303030303030303,
2118
+ 2.304414149172212
2119
+ ]
2120
+ },
2121
+ "count": {
2122
+ "items": 200,
2123
+ "gain": 7.000000000000001,
2124
+ "ci95": [
2125
+ 0.529030910609858,
2126
+ 13.588191699604737
2127
+ ]
2128
+ },
2129
+ "behavior": {
2130
+ "gain": 3.2500000000000004,
2131
+ "ci95": [
2132
+ -0.11554094741067852,
2133
+ 6.813306330633063
2134
+ ]
2135
+ }
2136
+ },
2137
+ {
2138
+ "layer": 7,
2139
+ "seed": 13,
2140
+ "color": {
2141
+ "items": 200,
2142
+ "gain": -2.0,
2143
+ "ci95": [
2144
+ -5.26379048582996,
2145
+ 0.970873786407767
2146
+ ]
2147
+ },
2148
+ "count": {
2149
+ "items": 200,
2150
+ "gain": 7.000000000000001,
2151
+ "ci95": [
2152
+ 0.0,
2153
+ 14.57286432160804
2154
+ ]
2155
+ },
2156
+ "behavior": {
2157
+ "gain": 2.5000000000000004,
2158
+ "ci95": [
2159
+ -1.2755678320738824,
2160
+ 6.4819444444444425
2161
+ ]
2162
+ }
2163
+ },
2164
+ {
2165
+ "layer": 9,
2166
+ "seed": 13,
2167
+ "color": {
2168
+ "items": 200,
2169
+ "gain": -1.0,
2170
+ "ci95": [
2171
+ -4.306737670224512,
2172
+ 2.0204583910167653
2173
+ ]
2174
+ },
2175
+ "count": {
2176
+ "items": 200,
2177
+ "gain": 4.5,
2178
+ "ci95": [
2179
+ -1.7275461106655972,
2180
+ 11.458956339712916
2181
+ ]
2182
+ },
2183
+ "behavior": {
2184
+ "gain": 1.75,
2185
+ "ci95": [
2186
+ -1.7086297740999141,
2187
+ 5.441599359855966
2188
+ ]
2189
+ }
2190
+ },
2191
+ {
2192
+ "layer": 11,
2193
+ "seed": 13,
2194
+ "color": {
2195
+ "items": 200,
2196
+ "gain": -2.0,
2197
+ "ci95": [
2198
+ -5.208938953488372,
2199
+ 0.9526098901098888
2200
+ ]
2201
+ },
2202
+ "count": {
2203
+ "items": 200,
2204
+ "gain": 4.0,
2205
+ "ci95": [
2206
+ -2.9275501622011295,
2207
+ 11.822792520790296
2208
+ ]
2209
+ },
2210
+ "behavior": {
2211
+ "gain": 1.0,
2212
+ "ci95": [
2213
+ -2.9426614935162325,
2214
+ 5.126663050541317
2215
+ ]
2216
+ }
2217
+ },
2218
+ {
2219
+ "layer": 16,
2220
+ "seed": 13,
2221
+ "color": {
2222
+ "items": 200,
2223
+ "gain": -1.5,
2224
+ "ci95": [
2225
+ -4.7368421052631575,
2226
+ 1.087105013067236
2227
+ ]
2228
+ },
2229
+ "count": {
2230
+ "items": 200,
2231
+ "gain": 6.0,
2232
+ "ci95": [
2233
+ 0.0,
2234
+ 13.147589239138524
2235
+ ]
2236
+ },
2237
+ "behavior": {
2238
+ "gain": 2.25,
2239
+ "ci95": [
2240
+ -1.1105623381637695,
2241
+ 6.108861758763935
2242
+ ]
2243
+ }
2244
+ },
2245
+ {
2246
+ "layer": 20,
2247
+ "seed": 13,
2248
+ "color": {
2249
+ "items": 200,
2250
+ "gain": -1.5,
2251
+ "ci95": [
2252
+ -4.7368421052631575,
2253
+ 1.087105013067236
2254
+ ]
2255
+ },
2256
+ "count": {
2257
+ "items": 200,
2258
+ "gain": 4.5,
2259
+ "ci95": [
2260
+ -1.5465920623898297,
2261
+ 11.283505154639169
2262
+ ]
2263
+ },
2264
+ "behavior": {
2265
+ "gain": 1.5,
2266
+ "ci95": [
2267
+ -1.9800744163479962,
2268
+ 5.112937480623073
2269
+ ]
2270
+ }
2271
+ },
2272
+ {
2273
+ "layer": 25,
2274
+ "seed": 13,
2275
+ "color": {
2276
+ "items": 200,
2277
+ "gain": -1.0,
2278
+ "ci95": [
2279
+ -4.2867445054945055,
2280
+ 2.000251256281406
2281
+ ]
2282
+ },
2283
+ "count": {
2284
+ "items": 200,
2285
+ "gain": 5.5,
2286
+ "ci95": [
2287
+ -0.9902221565440126,
2288
+ 12.386568902410033
2289
+ ]
2290
+ },
2291
+ "behavior": {
2292
+ "gain": 2.25,
2293
+ "ci95": [
2294
+ -1.2933704860056685,
2295
+ 5.923184095302921
2296
+ ]
2297
+ }
2298
+ },
2299
+ {
2300
+ "layer": 29,
2301
+ "seed": 13,
2302
+ "color": {
2303
+ "items": 200,
2304
+ "gain": -0.5,
2305
+ "ci95": [
2306
+ -3.4482758620689653,
2307
+ 1.9233091787439602
2308
+ ]
2309
+ },
2310
+ "count": {
2311
+ "items": 200,
2312
+ "gain": 3.0,
2313
+ "ci95": [
2314
+ -3.8109040717736367,
2315
+ 9.944823696161523
2316
+ ]
2317
+ },
2318
+ "behavior": {
2319
+ "gain": 1.25,
2320
+ "ci95": [
2321
+ -2.252752299343037,
2322
+ 5.000559454785878
2323
+ ]
2324
+ }
2325
+ },
2326
+ {
2327
+ "layer": 34,
2328
+ "seed": 13,
2329
+ "color": {
2330
+ "items": 200,
2331
+ "gain": -1.5,
2332
+ "ci95": [
2333
+ -5.377403464166534,
2334
+ 2.5255729887709566
2335
+ ]
2336
+ },
2337
+ "count": {
2338
+ "items": 200,
2339
+ "gain": 7.000000000000001,
2340
+ "ci95": [
2341
+ 0.0,
2342
+ 14.436242618356504
2343
+ ]
2344
+ },
2345
+ "behavior": {
2346
+ "gain": 2.7500000000000004,
2347
+ "ci95": [
2348
+ -1.1416853126282702,
2349
+ 6.977125546464398
2350
+ ]
2351
+ }
2352
+ }
2353
+ ],
2354
+ "layers": [
2355
+ 1,
2356
+ 2,
2357
+ 7,
2358
+ 9,
2359
+ 11,
2360
+ 16,
2361
+ 20,
2362
+ 25,
2363
+ 29,
2364
+ 34
2365
+ ],
2366
+ "layer_gain": [
2367
+ 2.75,
2368
+ 3.2500000000000004,
2369
+ 2.5000000000000004,
2370
+ 1.75,
2371
+ 1.0,
2372
+ 2.25,
2373
+ 1.5,
2374
+ 2.25,
2375
+ 1.25,
2376
+ 2.7500000000000004
2377
+ ],
2378
+ "spearman_gamma_gain": -0.40729671426564323
2379
+ },
2380
+ "single_llava_removal": {
2381
+ "behavior": "removal",
2382
+ "runs": [
2383
+ {
2384
+ "layer": 1,
2385
+ "seed": 13,
2386
+ "color": {
2387
+ "items": 200,
2388
+ "gain": 13.0,
2389
+ "ci95": [
2390
+ 7.034296482412061,
2391
+ 18.906769470184614
2392
+ ]
2393
+ },
2394
+ "count": {
2395
+ "items": 200,
2396
+ "gain": 5.5,
2397
+ "ci95": [
2398
+ -0.5025760113699812,
2399
+ 11.483672248803826
2400
+ ]
2401
+ },
2402
+ "behavior": {
2403
+ "gain": 9.25,
2404
+ "ci95": [
2405
+ 5.142881597138968,
2406
+ 13.636871325705943
2407
+ ]
2408
+ }
2409
+ },
2410
+ {
2411
+ "layer": 2,
2412
+ "seed": 13,
2413
+ "color": {
2414
+ "items": 200,
2415
+ "gain": 12.5,
2416
+ "ci95": [
2417
+ 5.6980204596784905,
2418
+ 20.21276595744681
2419
+ ]
2420
+ },
2421
+ "count": {
2422
+ "items": 200,
2423
+ "gain": 0.0,
2424
+ "ci95": [
2425
+ -6.167400881057269,
2426
+ 6.598984771573604
2427
+ ]
2428
+ },
2429
+ "behavior": {
2430
+ "gain": 6.25,
2431
+ "ci95": [
2432
+ 1.6090638525344352,
2433
+ 11.239593392046283
2434
+ ]
2435
+ }
2436
+ },
2437
+ {
2438
+ "layer": 6,
2439
+ "seed": 13,
2440
+ "color": {
2441
+ "items": 200,
2442
+ "gain": 15.5,
2443
+ "ci95": [
2444
+ 9.895833333333334,
2445
+ 21.579674689235645
2446
+ ]
2447
+ },
2448
+ "count": {
2449
+ "items": 200,
2450
+ "gain": 5.0,
2451
+ "ci95": [
2452
+ -0.5559350333940493,
2453
+ 11.000229357798164
2454
+ ]
2455
+ },
2456
+ "behavior": {
2457
+ "gain": 10.25,
2458
+ "ci95": [
2459
+ 6.217265689904091,
2460
+ 14.493690890990543
2461
+ ]
2462
+ }
2463
+ },
2464
+ {
2465
+ "layer": 8,
2466
+ "seed": 13,
2467
+ "color": {
2468
+ "items": 200,
2469
+ "gain": 16.5,
2470
+ "ci95": [
2471
+ 10.144927536231885,
2472
+ 23.189264372553055
2473
+ ]
2474
+ },
2475
+ "count": {
2476
+ "items": 200,
2477
+ "gain": 2.0,
2478
+ "ci95": [
2479
+ -3.175146771037182,
2480
+ 7.620879120879111
2481
+ ]
2482
+ },
2483
+ "behavior": {
2484
+ "gain": 9.25,
2485
+ "ci95": [
2486
+ 5.1586331095872895,
2487
+ 13.585090843405434
2488
+ ]
2489
+ }
2490
+ },
2491
+ {
2492
+ "layer": 10,
2493
+ "seed": 13,
2494
+ "color": {
2495
+ "items": 200,
2496
+ "gain": 18.0,
2497
+ "ci95": [
2498
+ 11.650485436893204,
2499
+ 24.726301735647525
2500
+ ]
2501
+ },
2502
+ "count": {
2503
+ "items": 200,
2504
+ "gain": 4.5,
2505
+ "ci95": [
2506
+ -0.5524861878453039,
2507
+ 10.054112554112546
2508
+ ]
2509
+ },
2510
+ "behavior": {
2511
+ "gain": 11.25,
2512
+ "ci95": [
2513
+ 7.115203128628778,
2514
+ 15.669871857280869
2515
+ ]
2516
+ }
2517
+ },
2518
+ {
2519
+ "layer": 14,
2520
+ "seed": 13,
2521
+ "color": {
2522
+ "items": 200,
2523
+ "gain": 16.5,
2524
+ "ci95": [
2525
+ 10.76829268292683,
2526
+ 22.748815165876778
2527
+ ]
2528
+ },
2529
+ "count": {
2530
+ "items": 200,
2531
+ "gain": 4.5,
2532
+ "ci95": [
2533
+ -1.5001884422110552,
2534
+ 10.606060606060606
2535
+ ]
2536
+ },
2537
+ "behavior": {
2538
+ "gain": 10.5,
2539
+ "ci95": [
2540
+ 6.452646768436243,
2541
+ 14.983362424869926
2542
+ ]
2543
+ }
2544
+ },
2545
+ {
2546
+ "layer": 18,
2547
+ "seed": 13,
2548
+ "color": {
2549
+ "items": 200,
2550
+ "gain": 13.0,
2551
+ "ci95": [
2552
+ 6.498790322580645,
2553
+ 19.474937343358388
2554
+ ]
2555
+ },
2556
+ "count": {
2557
+ "items": 200,
2558
+ "gain": 4.5,
2559
+ "ci95": [
2560
+ -1.5001884422110552,
2561
+ 10.550516343183808
2562
+ ]
2563
+ },
2564
+ "behavior": {
2565
+ "gain": 8.75,
2566
+ "ci95": [
2567
+ 4.36626958475698,
2568
+ 13.041379126693098
2569
+ ]
2570
+ }
2571
+ },
2572
+ {
2573
+ "layer": 22,
2574
+ "seed": 13,
2575
+ "color": {
2576
+ "items": 200,
2577
+ "gain": 15.0,
2578
+ "ci95": [
2579
+ 8.808290155440414,
2580
+ 21.634615384615383
2581
+ ]
2582
+ },
2583
+ "count": {
2584
+ "items": 200,
2585
+ "gain": 6.0,
2586
+ "ci95": [
2587
+ 0.5126903553299492,
2588
+ 11.558699293569292
2589
+ ]
2590
+ },
2591
+ "behavior": {
2592
+ "gain": 10.5,
2593
+ "ci95": [
2594
+ 6.497500148216218,
2595
+ 14.925461449806292
2596
+ ]
2597
+ }
2598
+ },
2599
+ {
2600
+ "layer": 26,
2601
+ "seed": 13,
2602
+ "color": {
2603
+ "items": 200,
2604
+ "gain": 11.0,
2605
+ "ci95": [
2606
+ 5.290941697191697,
2607
+ 17.18836805555555
2608
+ ]
2609
+ },
2610
+ "count": {
2611
+ "items": 200,
2612
+ "gain": 2.0,
2613
+ "ci95": [
2614
+ -3.5175879396984926,
2615
+ 7.368939590355195
2616
+ ]
2617
+ },
2618
+ "behavior": {
2619
+ "gain": 6.5,
2620
+ "ci95": [
2621
+ 2.6321138211382116,
2622
+ 10.546899293754509
2623
+ ]
2624
+ }
2625
+ },
2626
+ {
2627
+ "layer": 30,
2628
+ "seed": 13,
2629
+ "color": {
2630
+ "items": 200,
2631
+ "gain": 15.0,
2632
+ "ci95": [
2633
+ 8.910631900886424,
2634
+ 21.31997738590709
2635
+ ]
2636
+ },
2637
+ "count": {
2638
+ "items": 200,
2639
+ "gain": 4.5,
2640
+ "ci95": [
2641
+ -0.975609756097561,
2642
+ 10.153331526292439
2643
+ ]
2644
+ },
2645
+ "behavior": {
2646
+ "gain": 9.75,
2647
+ "ci95": [
2648
+ 5.611576437951277,
2649
+ 14.282649948085417
2650
+ ]
2651
+ }
2652
+ }
2653
+ ],
2654
+ "layers": [
2655
+ 1,
2656
+ 2,
2657
+ 6,
2658
+ 8,
2659
+ 10,
2660
+ 14,
2661
+ 18,
2662
+ 22,
2663
+ 26,
2664
+ 30
2665
+ ],
2666
+ "layer_gain": [
2667
+ 9.25,
2668
+ 6.25,
2669
+ 10.25,
2670
+ 9.25,
2671
+ 11.25,
2672
+ 10.5,
2673
+ 8.75,
2674
+ 10.5,
2675
+ 6.5,
2676
+ 9.75
2677
+ ]
2678
+ },
2679
+ "single_llava_replacement": {
2680
+ "behavior": "replacement",
2681
+ "runs": [
2682
+ {
2683
+ "layer": 1,
2684
+ "seed": 13,
2685
+ "color": {
2686
+ "items": 200,
2687
+ "gain": 4.0,
2688
+ "ci95": [
2689
+ 0.5050505050505051,
2690
+ 7.651324085750309
2691
+ ]
2692
+ },
2693
+ "count": {
2694
+ "items": 200,
2695
+ "gain": -6.0,
2696
+ "ci95": [
2697
+ -12.935774495398656,
2698
+ 1.5469989744696897
2699
+ ]
2700
+ },
2701
+ "behavior": {
2702
+ "gain": -1.0,
2703
+ "ci95": [
2704
+ -4.976266438879213,
2705
+ 3.0546531505058674
2706
+ ]
2707
+ }
2708
+ },
2709
+ {
2710
+ "layer": 2,
2711
+ "seed": 13,
2712
+ "color": {
2713
+ "items": 200,
2714
+ "gain": 4.0,
2715
+ "ci95": [
2716
+ 0.48303372352285395,
2717
+ 7.870655446086479
2718
+ ]
2719
+ },
2720
+ "count": {
2721
+ "items": 200,
2722
+ "gain": -7.000000000000001,
2723
+ "ci95": [
2724
+ -14.354189669979142,
2725
+ 0.4904411764705869
2726
+ ]
2727
+ },
2728
+ "behavior": {
2729
+ "gain": -1.5000000000000004,
2730
+ "ci95": [
2731
+ -5.3814031440538095,
2732
+ 2.7080512007290194
2733
+ ]
2734
+ }
2735
+ },
2736
+ {
2737
+ "layer": 6,
2738
+ "seed": 13,
2739
+ "color": {
2740
+ "items": 200,
2741
+ "gain": 4.0,
2742
+ "ci95": [
2743
+ 0.4431818181818188,
2744
+ 7.805068597560974
2745
+ ]
2746
+ },
2747
+ "count": {
2748
+ "items": 200,
2749
+ "gain": -8.0,
2750
+ "ci95": [
2751
+ -15.657821672219576,
2752
+ 0.0
2753
+ ]
2754
+ },
2755
+ "behavior": {
2756
+ "gain": -2.0,
2757
+ "ci95": [
2758
+ -6.17859386804219,
2759
+ 2.296907461640508
2760
+ ]
2761
+ }
2762
+ },
2763
+ {
2764
+ "layer": 8,
2765
+ "seed": 13,
2766
+ "color": {
2767
+ "items": 200,
2768
+ "gain": 2.5,
2769
+ "ci95": [
2770
+ -1.442481884057971,
2771
+ 6.467661691542289
2772
+ ]
2773
+ },
2774
+ "count": {
2775
+ "items": 200,
2776
+ "gain": -9.5,
2777
+ "ci95": [
2778
+ -16.582469775474955,
2779
+ -2.0405573396871453
2780
+ ]
2781
+ },
2782
+ "behavior": {
2783
+ "gain": -3.5,
2784
+ "ci95": [
2785
+ -7.5131492665535005,
2786
+ 0.613791719025769
2787
+ ]
2788
+ }
2789
+ },
2790
+ {
2791
+ "layer": 10,
2792
+ "seed": 13,
2793
+ "color": {
2794
+ "items": 200,
2795
+ "gain": 5.0,
2796
+ "ci95": [
2797
+ 1.0927476835352816,
2798
+ 9.18381136238279
2799
+ ]
2800
+ },
2801
+ "count": {
2802
+ "items": 200,
2803
+ "gain": -4.0,
2804
+ "ci95": [
2805
+ -11.220714531088973,
2806
+ 3.5723490427098628
2807
+ ]
2808
+ },
2809
+ "behavior": {
2810
+ "gain": 0.5,
2811
+ "ci95": [
2812
+ -3.5706329606488394,
2813
+ 4.712367166174262
2814
+ ]
2815
+ }
2816
+ },
2817
+ {
2818
+ "layer": 14,
2819
+ "seed": 13,
2820
+ "color": {
2821
+ "items": 200,
2822
+ "gain": 3.0,
2823
+ "ci95": [
2824
+ -0.576704545454545,
2825
+ 7.027027027027027
2826
+ ]
2827
+ },
2828
+ "count": {
2829
+ "items": 200,
2830
+ "gain": -10.0,
2831
+ "ci95": [
2832
+ -17.55477620989968,
2833
+ -2.1273781380164376
2834
+ ]
2835
+ },
2836
+ "behavior": {
2837
+ "gain": -3.5,
2838
+ "ci95": [
2839
+ -7.797335793357239,
2840
+ 0.7975996096261714
2841
+ ]
2842
+ }
2843
+ },
2844
+ {
2845
+ "layer": 18,
2846
+ "seed": 13,
2847
+ "color": {
2848
+ "items": 200,
2849
+ "gain": 4.0,
2850
+ "ci95": [
2851
+ 0.0,
2852
+ 8.45853280765981
2853
+ ]
2854
+ },
2855
+ "count": {
2856
+ "items": 200,
2857
+ "gain": -7.000000000000001,
2858
+ "ci95": [
2859
+ -14.136257866624359,
2860
+ 0.5027048507845339
2861
+ ]
2862
+ },
2863
+ "behavior": {
2864
+ "gain": -1.5000000000000004,
2865
+ "ci95": [
2866
+ -5.48551157337036,
2867
+ 2.680202128574467
2868
+ ]
2869
+ }
2870
+ },
2871
+ {
2872
+ "layer": 22,
2873
+ "seed": 13,
2874
+ "color": {
2875
+ "items": 200,
2876
+ "gain": 1.5,
2877
+ "ci95": [
2878
+ -1.6044937036398137,
2879
+ 4.878048780487805
2880
+ ]
2881
+ },
2882
+ "count": {
2883
+ "items": 200,
2884
+ "gain": -6.5,
2885
+ "ci95": [
2886
+ -14.217436974789916,
2887
+ 1.4564882784750168
2888
+ ]
2889
+ },
2890
+ "behavior": {
2891
+ "gain": -2.5,
2892
+ "ci95": [
2893
+ -6.612527369599109,
2894
+ 1.8048049812029971
2895
+ ]
2896
+ }
2897
+ },
2898
+ {
2899
+ "layer": 26,
2900
+ "seed": 13,
2901
+ "color": {
2902
+ "items": 200,
2903
+ "gain": 3.0,
2904
+ "ci95": [
2905
+ 0.0,
2906
+ 6.011409506549136
2907
+ ]
2908
+ },
2909
+ "count": {
2910
+ "items": 200,
2911
+ "gain": -4.5,
2912
+ "ci95": [
2913
+ -11.708751793400287,
2914
+ 3.256578229064304
2915
+ ]
2916
+ },
2917
+ "behavior": {
2918
+ "gain": -0.75,
2919
+ "ci95": [
2920
+ -4.5458532695374805,
2921
+ 3.291276580833568
2922
+ ]
2923
+ }
2924
+ },
2925
+ {
2926
+ "layer": 30,
2927
+ "seed": 13,
2928
+ "color": {
2929
+ "items": 200,
2930
+ "gain": 2.0,
2931
+ "ci95": [
2932
+ -0.9952748818720468,
2933
+ 5.208333333333333
2934
+ ]
2935
+ },
2936
+ "count": {
2937
+ "items": 200,
2938
+ "gain": -3.5000000000000004,
2939
+ "ci95": [
2940
+ -9.43443161550737,
2941
+ 2.592044301888854
2942
+ ]
2943
+ },
2944
+ "behavior": {
2945
+ "gain": -0.7500000000000002,
2946
+ "ci95": [
2947
+ -3.9838911210774635,
2948
+ 2.644496880499118
2949
+ ]
2950
+ }
2951
+ }
2952
+ ],
2953
+ "layers": [
2954
+ 1,
2955
+ 2,
2956
+ 6,
2957
+ 8,
2958
+ 10,
2959
+ 14,
2960
+ 18,
2961
+ 22,
2962
+ 26,
2963
+ 30
2964
+ ],
2965
+ "layer_gain": [
2966
+ -1.0,
2967
+ -1.5000000000000004,
2968
+ -2.0,
2969
+ -3.5,
2970
+ 0.5,
2971
+ -3.5,
2972
+ -1.5000000000000004,
2973
+ -2.5,
2974
+ -0.75,
2975
+ -0.7500000000000002
2976
+ ]
2977
+ },
2978
+ "single_llava_context": {
2979
+ "behavior": "context",
2980
+ "runs": [
2981
+ {
2982
+ "layer": 1,
2983
+ "seed": 13,
2984
+ "color": {
2985
+ "items": 200,
2986
+ "gain": 3.0,
2987
+ "ci95": [
2988
+ -1.442481884057971,
2989
+ 7.575757575757576
2990
+ ]
2991
+ },
2992
+ "count": {
2993
+ "items": 200,
2994
+ "gain": 1.0,
2995
+ "ci95": [
2996
+ -5.855100324402259,
2997
+ 7.895042839657281
2998
+ ]
2999
+ },
3000
+ "behavior": {
3001
+ "gain": 2.0,
3002
+ "ci95": [
3003
+ -2.047191712236892,
3004
+ 6.097149013672297
3005
+ ]
3006
+ }
3007
+ },
3008
+ {
3009
+ "layer": 2,
3010
+ "seed": 13,
3011
+ "color": {
3012
+ "items": 200,
3013
+ "gain": 2.5,
3014
+ "ci95": [
3015
+ -2.128231525966598,
3016
+ 7.3040325901523575
3017
+ ]
3018
+ },
3019
+ "count": {
3020
+ "items": 200,
3021
+ "gain": 3.5000000000000004,
3022
+ "ci95": [
3023
+ -3.333732057416268,
3024
+ 10.606060606060606
3025
+ ]
3026
+ },
3027
+ "behavior": {
3028
+ "gain": 3.0,
3029
+ "ci95": [
3030
+ -1.022802887059845,
3031
+ 7.169356197249638
3032
+ ]
3033
+ }
3034
+ },
3035
+ {
3036
+ "layer": 6,
3037
+ "seed": 13,
3038
+ "color": {
3039
+ "items": 200,
3040
+ "gain": 3.0,
3041
+ "ci95": [
3042
+ -0.9757293161166906,
3043
+ 6.8631341425459045
3044
+ ]
3045
+ },
3046
+ "count": {
3047
+ "items": 200,
3048
+ "gain": 3.0,
3049
+ "ci95": [
3050
+ -4.433497536945813,
3051
+ 10.5277081592871
3052
+ ]
3053
+ },
3054
+ "behavior": {
3055
+ "gain": 3.0,
3056
+ "ci95": [
3057
+ -1.2488399525333325,
3058
+ 7.152922077865719
3059
+ ]
3060
+ }
3061
+ },
3062
+ {
3063
+ "layer": 8,
3064
+ "seed": 13,
3065
+ "color": {
3066
+ "items": 200,
3067
+ "gain": 2.0,
3068
+ "ci95": [
3069
+ -2.1052631578947367,
3070
+ 6.315789473684211
3071
+ ]
3072
+ },
3073
+ "count": {
3074
+ "items": 200,
3075
+ "gain": 5.5,
3076
+ "ci95": [
3077
+ -2.3006210165059646,
3078
+ 12.935774495398652
3079
+ ]
3080
+ },
3081
+ "behavior": {
3082
+ "gain": 3.75,
3083
+ "ci95": [
3084
+ -0.49279556650246303,
3085
+ 8.069760879156807
3086
+ ]
3087
+ }
3088
+ },
3089
+ {
3090
+ "layer": 10,
3091
+ "seed": 13,
3092
+ "color": {
3093
+ "items": 200,
3094
+ "gain": -1.0,
3095
+ "ci95": [
3096
+ -5.853658536585366,
3097
+ 3.773839877613461
3098
+ ]
3099
+ },
3100
+ "count": {
3101
+ "items": 200,
3102
+ "gain": 1.5,
3103
+ "ci95": [
3104
+ -5.670103092783505,
3105
+ 8.719072164948447
3106
+ ]
3107
+ },
3108
+ "behavior": {
3109
+ "gain": 0.25,
3110
+ "ci95": [
3111
+ -4.246823119811982,
3112
+ 4.598743639655034
3113
+ ]
3114
+ }
3115
+ },
3116
+ {
3117
+ "layer": 14,
3118
+ "seed": 13,
3119
+ "color": {
3120
+ "items": 200,
3121
+ "gain": 2.0,
3122
+ "ci95": [
3123
+ -3.1092130829015545,
3124
+ 7.143772893772889
3125
+ ]
3126
+ },
3127
+ "count": {
3128
+ "items": 200,
3129
+ "gain": 4.0,
3130
+ "ci95": [
3131
+ -2.5387218534426657,
3132
+ 10.68047850208044
3133
+ ]
3134
+ },
3135
+ "behavior": {
3136
+ "gain": 3.0,
3137
+ "ci95": [
3138
+ -1.25,
3139
+ 7.214662523777282
3140
+ ]
3141
+ }
3142
+ },
3143
+ {
3144
+ "layer": 18,
3145
+ "seed": 13,
3146
+ "color": {
3147
+ "items": 200,
3148
+ "gain": 1.0,
3149
+ "ci95": [
3150
+ -4.1894034363247545,
3151
+ 6.522012235088484
3152
+ ]
3153
+ },
3154
+ "count": {
3155
+ "items": 200,
3156
+ "gain": 5.0,
3157
+ "ci95": [
3158
+ -1.4087844603164006,
3159
+ 11.766154737757164
3160
+ ]
3161
+ },
3162
+ "behavior": {
3163
+ "gain": 3.0,
3164
+ "ci95": [
3165
+ -1.1368897306397308,
3166
+ 7.464407135382959
3167
+ ]
3168
+ }
3169
+ },
3170
+ {
3171
+ "layer": 22,
3172
+ "seed": 13,
3173
+ "color": {
3174
+ "items": 200,
3175
+ "gain": 1.0,
3176
+ "ci95": [
3177
+ -3.500439698492462,
3178
+ 5.613448209472462
3179
+ ]
3180
+ },
3181
+ "count": {
3182
+ "items": 200,
3183
+ "gain": 6.5,
3184
+ "ci95": [
3185
+ -0.5105379848274388,
3186
+ 14.215686274509803
3187
+ ]
3188
+ },
3189
+ "behavior": {
3190
+ "gain": 3.75,
3191
+ "ci95": [
3192
+ -0.4799292429242924,
3193
+ 8.190972176996683
3194
+ ]
3195
+ }
3196
+ },
3197
+ {
3198
+ "layer": 26,
3199
+ "seed": 13,
3200
+ "color": {
3201
+ "items": 200,
3202
+ "gain": 2.5,
3203
+ "ci95": [
3204
+ -1.9801980198019802,
3205
+ 6.967101114147562
3206
+ ]
3207
+ },
3208
+ "count": {
3209
+ "items": 200,
3210
+ "gain": 3.0,
3211
+ "ci95": [
3212
+ -3.535802184279342,
3213
+ 9.853307163382588
3214
+ ]
3215
+ },
3216
+ "behavior": {
3217
+ "gain": 2.75,
3218
+ "ci95": [
3219
+ -1.1970661826842977,
3220
+ 6.955142240616607
3221
+ ]
3222
+ }
3223
+ },
3224
+ {
3225
+ "layer": 30,
3226
+ "seed": 13,
3227
+ "color": {
3228
+ "items": 200,
3229
+ "gain": -0.5,
3230
+ "ci95": [
3231
+ -4.592425431711146,
3232
+ 3.5714285714285716
3233
+ ]
3234
+ },
3235
+ "count": {
3236
+ "items": 200,
3237
+ "gain": -1.0,
3238
+ "ci95": [
3239
+ -6.599886929535448,
3240
+ 4.712165365998222
3241
+ ]
3242
+ },
3243
+ "behavior": {
3244
+ "gain": -0.75,
3245
+ "ci95": [
3246
+ -4.25,
3247
+ 2.7513551953335282
3248
+ ]
3249
+ }
3250
+ }
3251
+ ],
3252
+ "layers": [
3253
+ 1,
3254
+ 2,
3255
+ 6,
3256
+ 8,
3257
+ 10,
3258
+ 14,
3259
+ 18,
3260
+ 22,
3261
+ 26,
3262
+ 30
3263
+ ],
3264
+ "layer_gain": [
3265
+ 2.0,
3266
+ 3.0,
3267
+ 3.0,
3268
+ 3.75,
3269
+ 0.25,
3270
+ 3.0,
3271
+ 3.0,
3272
+ 3.75,
3273
+ 2.75,
3274
+ -0.75
3275
+ ]
3276
+ }
3277
+ },
3278
+ "uploads_required": false,
3279
+ "llava_gamma_gain_correlation": "Not computed: this experiment has no LLaVA training-localization profile"
3280
+ }
reports/training/pilot_internvl_decision.json ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "lr": 5e-05,
3
+ "schedule": "cosine",
4
+ "epochs": 1,
5
+ "note": "best eligible mean gain +5.19; 4 setting(s) within 0.5",
6
+ "table": [
7
+ {
8
+ "lr": 2e-05,
9
+ "schedule": "constant",
10
+ "epochs": 1,
11
+ "mean_gain": 3.8518518518518547,
12
+ "eligible": true
13
+ },
14
+ {
15
+ "lr": 2e-05,
16
+ "schedule": "constant",
17
+ "epochs": 2,
18
+ "mean_gain": 4.296296296296299,
19
+ "eligible": true
20
+ },
21
+ {
22
+ "lr": 5e-05,
23
+ "schedule": "cosine",
24
+ "epochs": 1,
25
+ "mean_gain": 5.185185185185189,
26
+ "eligible": true
27
+ },
28
+ {
29
+ "lr": 5e-05,
30
+ "schedule": "cosine",
31
+ "epochs": 2,
32
+ "mean_gain": 4.8148148148148175,
33
+ "eligible": true
34
+ },
35
+ {
36
+ "lr": 0.00012,
37
+ "schedule": "cosine",
38
+ "epochs": 1,
39
+ "mean_gain": 5.185185185185187,
40
+ "eligible": true
41
+ },
42
+ {
43
+ "lr": 0.00012,
44
+ "schedule": "cosine",
45
+ "epochs": 2,
46
+ "mean_gain": 5.1851851851851904,
47
+ "eligible": true
48
+ }
49
+ ]
50
+ }
reports/training/pilot_llava_decision.json ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "lr": 2e-05,
3
+ "schedule": "constant",
4
+ "epochs": 2,
5
+ "note": "best eligible mean gain +2.59; 2 setting(s) within 0.5",
6
+ "table": [
7
+ {
8
+ "lr": 2e-05,
9
+ "schedule": "constant",
10
+ "epochs": 1,
11
+ "mean_gain": 0.4444444444444486,
12
+ "eligible": true
13
+ },
14
+ {
15
+ "lr": 2e-05,
16
+ "schedule": "constant",
17
+ "epochs": 2,
18
+ "mean_gain": 2.296296296296301,
19
+ "eligible": true
20
+ },
21
+ {
22
+ "lr": 5e-05,
23
+ "schedule": "cosine",
24
+ "epochs": 1,
25
+ "mean_gain": 1.4814814814814878,
26
+ "eligible": true
27
+ },
28
+ {
29
+ "lr": 5e-05,
30
+ "schedule": "cosine",
31
+ "epochs": 2,
32
+ "mean_gain": 2.0740740740740784,
33
+ "eligible": true
34
+ },
35
+ {
36
+ "lr": 0.00012,
37
+ "schedule": "cosine",
38
+ "epochs": 1,
39
+ "mean_gain": 1.6296296296296306,
40
+ "eligible": true
41
+ },
42
+ {
43
+ "lr": 0.00012,
44
+ "schedule": "cosine",
45
+ "epochs": 2,
46
+ "mean_gain": 2.5925925925925957,
47
+ "eligible": true
48
+ }
49
+ ]
50
+ }