Prompt Injection Detection on MIPIAD aggregate over English and Bangla (test)
89Accuracy (Acc)Hybrid (XLPID+TF-IDF)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Hybrid (XLPID+TF-IDF)Components=XLPID + TF-IDF2026.05 | 89 | 95.14 | 89.15 | 92.05 | 93.46 | 97.05 | 94.79 | |
| Boosting EnsembleMethod Type=Meta-ensemble2026.05 | 88.29 | 96.81 | 86.45 | 91.34 | 93.78 | 97.34 | 94.79 | |
| XLPID (Qwen2.5-1.5B)Backbone=Qwen2.5-1.5B, Training Setting=LoRA-fine-tuned2026.05 | 83.39 | 82.23 | 97.9 | 89.39 | 90.74 | 96.33 | 93.22 | |
| mBERTBackbone=mBERT2026.05 | 78.46 | 79.26 | 94.6 | 86.25 | 84.8 | 93.95 | 99.6 | |
| XLM-RoBERTaBackbone=XLM-RoBERTa2026.05 | 78.07 | 80.66 | 91.15 | 85.59 | 85.16 | 94.14 | 98.42 | |
| Stacking Ensemble†Method Type=Meta-ensemble2026.05 | 76.61 | 95.9 | 70.25 | 81.1 | 92.76 | 96.73 | 99.47 | |
| Qwen2.5-1.5B (no LoRA)Backbone=Qwen2.5-1.5B, Training Setting=no LoRA2026.05 | 76.5 | 77.32 | 94.95 | 85.23 | 81.6 | 92.68 | 97.33 | |
| XLM-RoBERTa-largeBackbone=XLM-RoBERTa-large2026.05 | 72.36 | 72.55 | 98.6 | 83.59 | 79.99 | 92.31 | 99.92 | |
| TF-IDF + SVMClassifier=SVM, Features=TF-IDF2026.05 | 72.11 | 94.33 | 64.85 | 76.86 | 84.16 | 93.11 | 98.19 | |
| DeBERTa-v3-large (EN-only)Backbone=DeBERTa-v3-large, Training Setting=English-only2026.05 | 71.43 | 71.43 | 100 | 83.33 | 52.25 | 72.53 | 99.81 | |
| TF-IDF + LRClassifier=Logistic Regression, Features=TF-IDF2026.05 | 68.21 | 93.5 | 59.65 | 72.83 | 83.42 | 92.55 | 97.57 |