Transformers
Safetensors
PEFT
Generated from Trainer
trl
sft
lora
safety-classifier
constitutional-ai
hf_jobs
trackio
Instructions to use imadreamerboy/constitutional-safety-classifier with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use imadreamerboy/constitutional-safety-classifier with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("imadreamerboy/constitutional-safety-classifier", device_map="auto") - PEFT
How to use imadreamerboy/constitutional-safety-classifier with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Add paper-aligned constitutional classifier evaluation results
Browse files- 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 |
+
}
|