Dataset: alirezaaminzadeh/secembed-pairs
Loss: MultipleNegativesRankingLoss (bi-encoder) / Binary cross-entropy (reranker)
Hard negatives: sibling ATT&CK techniques, same-CWE CVEs, unrelated Sigma rules
Hardware: Hugging Face ZeroGPU

Evaluation (CyberSec Retrieval Benchmark sample)

json

{
  "note": "pair classification trained; use demo Space for retrieval@k with bi-encoder cascade"
}

Usage

python

from sentence_transformers import CrossEncoder

model = CrossEncoder("alirezaaminzadeh/SecReranker")
print(model.predict([["detect powershell encoded command", "T1059.001 PowerShell..."]]))

Intended use

RAG over security knowledge bases, threat-intelligence search, SOC alert enrichment, CVE/CWE similarity, Sigma rule retrieval, and ATT&CK technique mapping.