Toxicity Classification on Toxicity
90.4Original AccuracyAdvDemo + CW
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| AdvDemo + CWRecipe Code=p10_CWmessage1_CWpos22026.01 | 90.4 | 3.8 | 61.6 | 80 | 63.6 | |
| AdvDemo + Random TemplateRecipe Code=p10_length102026.01 | 90.4 | 7 | 9.8 | 0 | 94.8 | |
| AdvDemo + CW + Random TemplateVariant=Toxicity I, Recipe Code=p10_15_10_length10_CWmessage1_CWpos02026.01 | 90.4 | 11.6 | 57.2 | 9,360 | 94.8 | |
| AdvDemo + CW + Random TemplateVariant=Toxicity II, Recipe Code=p10_length10_CWmessage1_CWpos12026.01 | 90.4 | 4.2 | 53.4 | 0 | 94.8 | |
| Llama-3.3-70B-Instruct-FP8N (#rows)=1,000, K (#candidate labels)=2, Decoding Strategy=greedy2026.05 | 80.1 | — | — | — | — | |
| gpt-oss-120BN (#rows)=1,000, K (#candidate labels)=2, Decoding Strategy=prefill-only logit scoring2026.05 | 78.7 | — | — | — | — |