Harm Classification on FHM
74.1AccuracyMuPHIRM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MuPHIRMTraining Paradigm=GRPO-based, Warmup=Yes2026.05 | 74.1 | 74.1 | |
| MuPHIRM w/o warmupTraining Paradigm=GRPO-based, Warmup=No2026.05 | 65.7 | 65.4 | |
| Label + RationaleTraining Paradigm=SFT2026.05 | 62.8 | 61.9 | |
| Zero-shot + DecompositionTraining Paradigm=Inference-only2026.05 | 57.2 | 49.8 | |
| Zero-shotTraining Paradigm=Inference-only2026.05 | 56 | 53 | |
| Zero-shot + CoTTraining Paradigm=Inference-only2026.05 | 52.7 | 43.3 | |
| Random2026.05 | 50 | 50 |