Threat Classification on Synthetic benchmark 11k samples (test)
24.4AccuracySALM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SALMMethod=Semantic retrieval via contrastive pretraining2026.03 | 24.4 | 14.3 | |
| FT CodeBERT + MLPBackbone=CodeBERT, Classifier=Multi-layer perceptron (MLP), Training Mode=Fine-tuned2026.03 | 20.6 | 13.2 | |
| TF-IDF + RFFeatures=sparse TF-IDF vectors, Classifier=Random Forest2026.03 | 16 | 9 | |
| Embedding SimilarityMethod=RAG-style / k-nearest-neighbor majority voting, Backbone=Sentence-BERT (all-MiniLM-L6-v2), Search Index=HNSW2026.03 | 8.1 | 4.3 |