Adversarial Intent Classification on Experiment 3 Conversations (test)
42.5Precisionllama-guard-4-12b (LG)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| llama-guard-4-12b (LG)k (turns removed)=0, Prefix (retained turns)=6, Encoder=lexical encoder ϕL2026.06 | 42.5 | 24.2 | 30.8 | 57.5 | |
| GBk (turns removed)=3, Prefix (retained turns)=3, Encoder=lexical encoder ϕL2026.06 | 31.9 | 67.4 | 43.3 | 65.1 | |
| LRk (turns removed)=0, Prefix (retained turns)=6, Encoder=lexical encoder ϕL2026.06 | 31.8 | 72.1 | 44.2 | 69.7 | |
| GBk (turns removed)=0, Prefix (retained turns)=6, Encoder=lexical encoder ϕL2026.06 | 31.7 | 70.5 | 43.8 | 68.2 | |
| LRk (turns removed)=1, Prefix (retained turns)=5, Encoder=lexical encoder ϕL2026.06 | 31.6 | 63.6 | 42.2 | 66.8 | |
| GBk (turns removed)=1, Prefix (retained turns)=5, Encoder=lexical encoder ϕL2026.06 | 31.5 | 75.8 | 44.5 | 65.9 | |
| LRk (turns removed)=2, Prefix (retained turns)=4, Encoder=lexical encoder ϕL2026.06 | 31.5 | 60.3 | 41.4 | 64.8 | |
| LRk (turns removed)=4, Prefix (retained turns)=2, Encoder=lexical encoder ϕL2026.06 | 30.7 | 59.3 | 40.5 | 60.9 | |
| llama-guard-4-12b (LG)k (turns removed)=1, Prefix (retained turns)=5, Encoder=lexical encoder ϕL2026.06 | 28.8 | 15.9 | 20.5 | 52.2 | |
| GBk (turns removed)=2, Prefix (retained turns)=4, Encoder=lexical encoder ϕL2026.06 | 28.6 | 60.3 | 38.8 | 62.2 | |
| llama-guard-4-12b (LG)k (turns removed)=3, Prefix (retained turns)=3, Encoder=lexical encoder ϕL2026.06 | 28 | 10.4 | 15.1 | 51.2 | |
| GBk (turns removed)=4, Prefix (retained turns)=2, Encoder=lexical encoder ϕL2026.06 | 27.5 | 55 | 36.7 | 58.6 | |
| LRk (turns removed)=3, Prefix (retained turns)=3, Encoder=lexical encoder ϕL2026.06 | 26.6 | 59.3 | 36.7 | 63 | |
| llama-guard-4-12b (LG)k (turns removed)=2, Prefix (retained turns)=4, Encoder=lexical encoder ϕL2026.06 | 21.1 | 8.8 | 12.4 | 49.4 | |
| llama-guard-4-12b (LG)k (turns removed)=4, Prefix (retained turns)=2, Encoder=lexical encoder ϕL2026.06 | 16.7 | 4.3 | 6.8 | 48.8 |