imadreamerboy commited on
Commit
00aadf0
·
verified ·
1 Parent(s): 0ae551a

Add paper-aligned constitutional classifier evaluation results

Browse files
Files changed (1) hide show
  1. constitutional_eval_results.json +253 -0
constitutional_eval_results.json ADDED
@@ -0,0 +1,253 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "imadreamerboy/constitutional-safety-classifier",
3
+ "base_model": "Qwen/Qwen3-1.7B",
4
+ "paper_alignment": {
5
+ "paper": "Constitutional Classifiers (arXiv:2501.18837)",
6
+ "note": "Exact Anthropic jailbreaks and Claude.ai traffic are not public. Public proxies evaluate held-out accuracy, harmful recall/ASR-proxy, and over-refusal/FPR per Section 5."
7
+ },
8
+ "datasets": {
9
+ "heldout_constitutional_classifier_data": {
10
+ "n": 500,
11
+ "safe_n": 250,
12
+ "unsafe_n": 250,
13
+ "accuracy": 0.802,
14
+ "unsafe_precision": 0.7381703470031545,
15
+ "unsafe_recall_tpr": 0.936,
16
+ "unsafe_f1": 0.8253968253968254,
17
+ "over_refusal_fpr": 0.332,
18
+ "missed_unsafe_fnr_asr_proxy": 0.064,
19
+ "mean_unsafe_score": 0.644300048828125,
20
+ "threshold": 0.5,
21
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
22
+ [
23
+ 167,
24
+ 83
25
+ ],
26
+ [
27
+ 16,
28
+ 234
29
+ ]
30
+ ],
31
+ "auroc": 0.929824,
32
+ "auprc": 0.9352743249330117,
33
+ "threshold_sweep": {
34
+ "max_tpr_at_fpr<=0.005": {
35
+ "threshold": 0.997,
36
+ "tpr": 0.22,
37
+ "fpr": 0.0
38
+ },
39
+ "max_tpr_at_fpr<=0.01": {
40
+ "threshold": 0.997,
41
+ "tpr": 0.22,
42
+ "fpr": 0.0
43
+ },
44
+ "max_tpr_at_fpr<=0.05": {
45
+ "threshold": 0.981,
46
+ "tpr": 0.728,
47
+ "fpr": 0.032
48
+ }
49
+ }
50
+ },
51
+ "toxicchat_toxicity": {
52
+ "n": 500,
53
+ "safe_n": 250,
54
+ "unsafe_n": 250,
55
+ "accuracy": 0.9,
56
+ "unsafe_precision": 0.8571428571428571,
57
+ "unsafe_recall_tpr": 0.96,
58
+ "unsafe_f1": 0.9056603773584906,
59
+ "over_refusal_fpr": 0.16,
60
+ "missed_unsafe_fnr_asr_proxy": 0.04,
61
+ "mean_unsafe_score": 0.609771484375,
62
+ "threshold": 0.5,
63
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
64
+ [
65
+ 210,
66
+ 40
67
+ ],
68
+ [
69
+ 10,
70
+ 240
71
+ ]
72
+ ],
73
+ "auroc": 0.973272,
74
+ "auprc": 0.9734237903869332
75
+ },
76
+ "toxicchat_jailbreak_subset": {
77
+ "n": 66,
78
+ "safe_n": 0,
79
+ "unsafe_n": 66,
80
+ "accuracy": 0.9848484848484849,
81
+ "unsafe_precision": 1.0,
82
+ "unsafe_recall_tpr": 0.9848484848484849,
83
+ "unsafe_f1": 0.9923664122137404,
84
+ "over_refusal_fpr": 0.0,
85
+ "missed_unsafe_fnr_asr_proxy": 0.015151515151515152,
86
+ "mean_unsafe_score": 0.9376479640151515,
87
+ "threshold": 0.5,
88
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
89
+ [
90
+ 0,
91
+ 0
92
+ ],
93
+ [
94
+ 1,
95
+ 65
96
+ ]
97
+ ],
98
+ "auroc": null,
99
+ "auprc": null
100
+ },
101
+ "aegis2_prompt_only": {
102
+ "n": 500,
103
+ "safe_n": 250,
104
+ "unsafe_n": 250,
105
+ "accuracy": 0.752,
106
+ "unsafe_precision": 0.6769662921348315,
107
+ "unsafe_recall_tpr": 0.964,
108
+ "unsafe_f1": 0.7953795379537953,
109
+ "over_refusal_fpr": 0.46,
110
+ "missed_unsafe_fnr_asr_proxy": 0.036,
111
+ "mean_unsafe_score": 0.70212744140625,
112
+ "threshold": 0.5,
113
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
114
+ [
115
+ 135,
116
+ 115
117
+ ],
118
+ [
119
+ 9,
120
+ 241
121
+ ]
122
+ ],
123
+ "auroc": 0.9228399999999999,
124
+ "auprc": 0.9148293600951266
125
+ },
126
+ "aegis2_prompt_response": {
127
+ "n": 500,
128
+ "safe_n": 250,
129
+ "unsafe_n": 250,
130
+ "accuracy": 0.778,
131
+ "unsafe_precision": 0.698005698005698,
132
+ "unsafe_recall_tpr": 0.98,
133
+ "unsafe_f1": 0.8153078202995009,
134
+ "over_refusal_fpr": 0.424,
135
+ "missed_unsafe_fnr_asr_proxy": 0.02,
136
+ "mean_unsafe_score": 0.69241552734375,
137
+ "threshold": 0.5,
138
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
139
+ [
140
+ 144,
141
+ 106
142
+ ],
143
+ [
144
+ 5,
145
+ 245
146
+ ]
147
+ ],
148
+ "auroc": 0.93728,
149
+ "auprc": 0.9249390018989783
150
+ },
151
+ "beavertails_prompt_response": {
152
+ "n": 500,
153
+ "safe_n": 250,
154
+ "unsafe_n": 250,
155
+ "accuracy": 0.814,
156
+ "unsafe_precision": 0.7460815047021944,
157
+ "unsafe_recall_tpr": 0.952,
158
+ "unsafe_f1": 0.836555360281195,
159
+ "over_refusal_fpr": 0.324,
160
+ "missed_unsafe_fnr_asr_proxy": 0.048,
161
+ "mean_unsafe_score": 0.652507080078125,
162
+ "threshold": 0.5,
163
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
164
+ [
165
+ 169,
166
+ 81
167
+ ],
168
+ [
169
+ 12,
170
+ 238
171
+ ]
172
+ ],
173
+ "auroc": 0.9336800000000001,
174
+ "auprc": 0.9187361971235515
175
+ },
176
+ "or_bench_hard_overrefusal_expected_safe": {
177
+ "n": 500,
178
+ "safe_n": 500,
179
+ "unsafe_n": 0,
180
+ "accuracy": 0.05,
181
+ "unsafe_precision": 0.0,
182
+ "unsafe_recall_tpr": 0.0,
183
+ "unsafe_f1": 0.0,
184
+ "over_refusal_fpr": 0.95,
185
+ "missed_unsafe_fnr_asr_proxy": 0.0,
186
+ "mean_unsafe_score": 0.890638671875,
187
+ "threshold": 0.5,
188
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
189
+ [
190
+ 25,
191
+ 475
192
+ ],
193
+ [
194
+ 0,
195
+ 0
196
+ ]
197
+ ],
198
+ "auroc": null,
199
+ "auprc": null
200
+ },
201
+ "or_bench_toxic_expected_unsafe": {
202
+ "n": 500,
203
+ "safe_n": 0,
204
+ "unsafe_n": 500,
205
+ "accuracy": 0.996,
206
+ "unsafe_precision": 1.0,
207
+ "unsafe_recall_tpr": 0.996,
208
+ "unsafe_f1": 0.9979959919839679,
209
+ "over_refusal_fpr": 0.0,
210
+ "missed_unsafe_fnr_asr_proxy": 0.004,
211
+ "mean_unsafe_score": 0.979162109375,
212
+ "threshold": 0.5,
213
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
214
+ [
215
+ 0,
216
+ 0
217
+ ],
218
+ [
219
+ 2,
220
+ 498
221
+ ]
222
+ ],
223
+ "auroc": null,
224
+ "auprc": null
225
+ },
226
+ "mmlu_chemistry_overrefusal_expected_safe": {
227
+ "n": 303,
228
+ "safe_n": 303,
229
+ "unsafe_n": 0,
230
+ "accuracy": 0.9966996699669967,
231
+ "unsafe_precision": 0.0,
232
+ "unsafe_recall_tpr": 0.0,
233
+ "unsafe_f1": 0.0,
234
+ "over_refusal_fpr": 0.0033003300330033004,
235
+ "missed_unsafe_fnr_asr_proxy": 0.0,
236
+ "mean_unsafe_score": 0.22647870255775576,
237
+ "threshold": 0.5,
238
+ "confusion_matrix_[[tn_fp],[fn_tp]]": [
239
+ [
240
+ 302,
241
+ 1
242
+ ],
243
+ [
244
+ 0,
245
+ 0
246
+ ]
247
+ ],
248
+ "auroc": null,
249
+ "auprc": null
250
+ }
251
+ },
252
+ "elapsed_seconds": 551.519150018692
253
+ }