Language model detoxification on RealToxicityPrompts (test)

91.3Distinct-1TruthFlow

Evaluation results

Method
TruthFlow2026.02	91.3	—	—	—	99.1	99.5
Linear-AcT2026.02	91.2	—	—	—	99.3	99.8
HPR2026.02	91.1	—	—	—	99.3	99.8
Original2026.02	91	—	—	—	99.2	99.7
CAA2026.02	91	—	—	—	99.1	99.6
HPR2026.02	91	—	—	—	99.1	99.6
Original2026.02	90.9	—	—	—	99.1	99.7
RE-Control2026.02	90.9	—	—	—	99.2	99.7
MiMiC2026.02	90.9	—	—	—	99.1	99.6
MiMiC2026.02	90.8	—	—	—	99.2	99.8
Linear-AcT2026.02	90.7	—	—	—	99.1	99.7
CAA2026.02	90.7	—	—	—	99.1	99.6
Linear-AcT2026.02	90.7	—	—	—	99.1	99.7
CAA2026.02	90.6	—	—	—	99.1	99.7
MiMiC2026.02	90.6	—	—	—	99.1	99.7
RepE2026.02	90.6	—	—	—	99.1	99.7
ITI2026.02	90.6	—	—	—	99.1	99.7
ODESTEER2026.02	90.6	—	—	—	99.2	99.7
Original2026.02	90.5	—	—	—	99.1	99.7
ODESTEER2026.02	90.5	—	—	—	99.3	99.8
TruthFlow2026.02	90.5	—	—	—	99.2	99.8
ODESTEER2026.02	90.5	—	—	—	99.3	99.8
RE-Control2026.02	90.2	—	—	—	98.8	99.5
ITI2026.02	90.1	—	—	—	98.9	99.6
RE-Control2026.02	90.1	—	—	—	98.9	99.6
ITI2026.02	89.4	—	—	—	98.4	99.3
RepE2026.02	87.8	—	—	—	96.1	97.6
HPR2026.02	87.1	—	—	—	97.5	98.8
TruthFlow2026.02	86.6	—	—	—	97.7	98.9
Original2026.02	81	—	—	—	94.8	97.2
CAA2026.02	81	—	—	—	95	97.4
Linear-AcT2026.02	81	—	—	—	94.9	97.2
RE-Control2026.02	80.2	—	—	—	94.1	96.4
MiMiC2026.02	80.1	—	—	—	94.1	96.7
ODESTEER2026.02	79.8	—	—	—	94.4	96.9
RepE2026.02	79.7	—	—	—	94	96.6
ITI2026.02	79.6	—	—	—	93.5	96
RepE2026.02	77.4	—	—	—	96.9	99.4
TruthFlow2026.02	76.9	—	—	—	91	94.2
HPR2026.02	76.8	—	—	—	91.9	95
GeDi2023.05	62	0.363	0.217	60.03	84	83
GPT-2 (original)2023.05	58	0.527	0.52	25.45	85	85
PPLM (10%)2023.05	58	0.52	0.518	32.58	86	86
DExperts-large2023.05	58	0.314	0.128	32.41	84	84
PromptT52023.05	58	0.32	0.172	55.1	76	70
DAPT2023.05	57	0.428	0.36	31.21	84	84
DExperts-medium2023.05	57	0.307	0.125	32.51	84	84
DExperts-base2023.05	56	0.302	0.118	38.2	82	83
LM-Steer-medium2023.05	56	0.215	0.059	43.56	83	84
MuCoLa2023.05	55	0.308	0.088	29.92	82	83
LM-Steer-large2023.05	55	0.249	0.089	28.26	84	84
LM-Steer-base2023.05	54	0.296	0.129	36.87	86	86
LORA2023.05	53	0.365	0.21	21.11	85	86
Soft-Blacklist2023.05	53	0.27	0.154	18.28	81	83