Toxicity Classification on COLD (test)
94.33AccuracyCOLD
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| COLDBackbone=Qwen3-8B, Training Source=COLD2026.05 | 94.33 | — | — | — | — | |
| MixedBackbone=Qwen3-8B, Training Source=Mixed2026.05 | 90.8 | — | — | — | — | |
| CITDBackbone=Qwen3-8B, Training Source=CITD (CITA-generated + public)2026.05 | 88.8 | — | — | — | — | |
| RoBERTa + ToxiTraceTraining Strategy=ToxiTrace, Backbone=Chinese-RoBERTa-wwm-ext2026.04 | 83.84 | 86.19 | 76.14 | 80.85 | 83.68 | |
| MacBERT + ToxiTraceTraining Strategy=ToxiTrace, Backbone=Chinese-MacBERT-base2026.04 | 83.22 | 86.19 | 75.1 | 80.27 | 83.13 | |
| LLaMA3.1-8B (SFT)Training Strategy=SFT, Adaptation=LoRA2026.04 | 83 | 87.13 | 74.37 | 80.19 | 82.19 | |
| Qwen3-8B + ToxiTraceTraining Strategy=ToxiTrace, Adaptation=LoRA2026.04 | 82.96 | 82.49 | 76.38 | 79.41 | 82.39 | |
| RoBERTaBackbone=Chinese-RoBERTa-wwm-ext2026.04 | 82.68 | 86.37 | 74.42 | 79.8 | 82.56 | |
| LLaMA2-7B (SFT)Training Strategy=SFT, Adaptation=LoRA2026.04 | 82.43 | 86.28 | 73.78 | 79.54 | 82.46 | |
| Qwen3-8B (SFT)Training Strategy=SFT, Adaptation=LoRA2026.04 | 82.25 | 82.68 | 75.02 | 78.66 | 81.87 | |
| LLaMA3.1-8B + ToxiTraceTraining Strategy=ToxiTrace, Adaptation=LoRA2026.04 | 82.15 | 88.42 | 72.52 | 79.68 | 81.89 | |
| ToxiCNBackbone=Qwen3-8B, Training Source=ToxiCN2026.05 | 81 | — | — | — | — | |
| SCCDBackbone=Qwen3-8B, Training Source=SCCD2026.05 | 78.87 | — | — | — | — | |
| GPT-4oTraining Strategy=Zero-shot2026.04 | 74.49 | 67.79 | 66.98 | 67.38 | 73.22 | |
| DeepSeek-V3.2Training Strategy=Zero-shot2026.04 | 74.23 | 59.23 | 70.91 | 64.55 | 72.51 | |
| Qwen3-8BTraining Strategy=Zero-shot2026.04 | 74.03 | 59.9 | 70.15 | 64.62 | 72.33 | |
| LLaMA3.1-8BTraining Strategy=Zero-shot2026.04 | 73.91 | 51.46 | 74.58 | 60.9 | 72 | |
| CNTPBackbone=Qwen3-8B, Training Source=CNTP2026.05 | 70.47 | — | — | — | — | |
| SWSRBackbone=Qwen3-8B, Training Source=SWSR2026.05 | 47.87 | — | — | — | — |