{ "model": "imadreamerboy/constitutional-safety-classifier", "base_model": "Qwen/Qwen3-1.7B", "paper_alignment": { "paper": "Constitutional Classifiers (arXiv:2501.18837)", "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." }, "datasets": { "heldout_constitutional_classifier_data": { "n": 500, "safe_n": 250, "unsafe_n": 250, "accuracy": 0.802, "unsafe_precision": 0.7381703470031545, "unsafe_recall_tpr": 0.936, "unsafe_f1": 0.8253968253968254, "over_refusal_fpr": 0.332, "missed_unsafe_fnr_asr_proxy": 0.064, "mean_unsafe_score": 0.644300048828125, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 167, 83 ], [ 16, 234 ] ], "auroc": 0.929824, "auprc": 0.9352743249330117, "threshold_sweep": { "max_tpr_at_fpr<=0.005": { "threshold": 0.997, "tpr": 0.22, "fpr": 0.0 }, "max_tpr_at_fpr<=0.01": { "threshold": 0.997, "tpr": 0.22, "fpr": 0.0 }, "max_tpr_at_fpr<=0.05": { "threshold": 0.981, "tpr": 0.728, "fpr": 0.032 } } }, "toxicchat_toxicity": { "n": 500, "safe_n": 250, "unsafe_n": 250, "accuracy": 0.9, "unsafe_precision": 0.8571428571428571, "unsafe_recall_tpr": 0.96, "unsafe_f1": 0.9056603773584906, "over_refusal_fpr": 0.16, "missed_unsafe_fnr_asr_proxy": 0.04, "mean_unsafe_score": 0.609771484375, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 210, 40 ], [ 10, 240 ] ], "auroc": 0.973272, "auprc": 0.9734237903869332 }, "toxicchat_jailbreak_subset": { "n": 66, "safe_n": 0, "unsafe_n": 66, "accuracy": 0.9848484848484849, "unsafe_precision": 1.0, "unsafe_recall_tpr": 0.9848484848484849, "unsafe_f1": 0.9923664122137404, "over_refusal_fpr": 0.0, "missed_unsafe_fnr_asr_proxy": 0.015151515151515152, "mean_unsafe_score": 0.9376479640151515, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 0, 0 ], [ 1, 65 ] ], "auroc": null, "auprc": null }, "aegis2_prompt_only": { "n": 500, "safe_n": 250, "unsafe_n": 250, "accuracy": 0.752, "unsafe_precision": 0.6769662921348315, "unsafe_recall_tpr": 0.964, "unsafe_f1": 0.7953795379537953, "over_refusal_fpr": 0.46, "missed_unsafe_fnr_asr_proxy": 0.036, "mean_unsafe_score": 0.70212744140625, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 135, 115 ], [ 9, 241 ] ], "auroc": 0.9228399999999999, "auprc": 0.9148293600951266 }, "aegis2_prompt_response": { "n": 500, "safe_n": 250, "unsafe_n": 250, "accuracy": 0.778, "unsafe_precision": 0.698005698005698, "unsafe_recall_tpr": 0.98, "unsafe_f1": 0.8153078202995009, "over_refusal_fpr": 0.424, "missed_unsafe_fnr_asr_proxy": 0.02, "mean_unsafe_score": 0.69241552734375, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 144, 106 ], [ 5, 245 ] ], "auroc": 0.93728, "auprc": 0.9249390018989783 }, "beavertails_prompt_response": { "n": 500, "safe_n": 250, "unsafe_n": 250, "accuracy": 0.814, "unsafe_precision": 0.7460815047021944, "unsafe_recall_tpr": 0.952, "unsafe_f1": 0.836555360281195, "over_refusal_fpr": 0.324, "missed_unsafe_fnr_asr_proxy": 0.048, "mean_unsafe_score": 0.652507080078125, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 169, 81 ], [ 12, 238 ] ], "auroc": 0.9336800000000001, "auprc": 0.9187361971235515 }, "or_bench_hard_overrefusal_expected_safe": { "n": 500, "safe_n": 500, "unsafe_n": 0, "accuracy": 0.05, "unsafe_precision": 0.0, "unsafe_recall_tpr": 0.0, "unsafe_f1": 0.0, "over_refusal_fpr": 0.95, "missed_unsafe_fnr_asr_proxy": 0.0, "mean_unsafe_score": 0.890638671875, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 25, 475 ], [ 0, 0 ] ], "auroc": null, "auprc": null }, "or_bench_toxic_expected_unsafe": { "n": 500, "safe_n": 0, "unsafe_n": 500, "accuracy": 0.996, "unsafe_precision": 1.0, "unsafe_recall_tpr": 0.996, "unsafe_f1": 0.9979959919839679, "over_refusal_fpr": 0.0, "missed_unsafe_fnr_asr_proxy": 0.004, "mean_unsafe_score": 0.979162109375, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 0, 0 ], [ 2, 498 ] ], "auroc": null, "auprc": null }, "mmlu_chemistry_overrefusal_expected_safe": { "n": 303, "safe_n": 303, "unsafe_n": 0, "accuracy": 0.9966996699669967, "unsafe_precision": 0.0, "unsafe_recall_tpr": 0.0, "unsafe_f1": 0.0, "over_refusal_fpr": 0.0033003300330033004, "missed_unsafe_fnr_asr_proxy": 0.0, "mean_unsafe_score": 0.22647870255775576, "threshold": 0.5, "confusion_matrix_[[tn_fp],[fn_tp]]": [ [ 302, 1 ], [ 0, 0 ] ], "auroc": null, "auprc": null } }, "elapsed_seconds": 551.519150018692 }