SecEmbed-base

Cybersecurity dense embedding model fine-tuned from nswerdotai/ModernBERT-base on secembed-pairs.

Trained at max_seq_length=256 for efficient retrieval latency; ModernBERT remains capable of longer contexts at inference when needed.

Evaluation (CyberSec Retrieval Benchmark sample)

json { "attack_retrieval": { "recall@5": 1.0, "recall@10": 1.0, "n": 80 }, "cve_similarity": { "recall@5": 0.9625, "recall@10": 0.9625, "n": 80 }, "cwe_retrieval": { "recall@5": 1.0, "recall@10": 1.0, "n": 80 }, "sigma_retrieval": { "recall@5": 0.8875, "recall@10": 0.9, "n": 80 }, "soc_playbook": { "recall@5": 1.0, "recall@10": 1.0, "n": 80 }, "threat_report_retrieval": { "recall@5": 0.975, "recall@10": 0.975, "n": 80 }, "macro_recall@5": 0.9708 }

Baseline macro Recall@5 (untrained ModernBERT): 0.1417 Final macro Recall@5: 0.9708

Usage

python from sentence_transformers import SentenceTransformer model = SentenceTransformer("alirezaaminzadeh/SecEmbed-base") model.max_seq_length = 512 emb = model.encode(["remote desktop credential attack"])

Related

• Dataset: secembed-pairs
• Benchmark: cybersec-retrieval-benchmark
• Reranker: SecReranker
• Demo: https://huggingface.co/spaces/alirezaaminzadeh/secembed