nyxspecter4/kinetigor-dpo-cybersec

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 1, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

nyxspecter4/kinetigor-dpo-cybersec is a 0.5 billion parameter language model, DPO-tuned from Mistral-7B-Instruct-v0.2, specifically designed for cyber-defense and security analyst reasoning. It excels in identifying MITRE ATT&CK TTPs, aligning with OWASP/NIST/SANS frameworks, and ideating detection rules, leveraging a specialized dataset covering 10 cybersecurity frameworks. With a 32768 token context length, it focuses on high accuracy in a narrow cybersecurity domain.

Loading preview...

kinetigor-dpo-cybersec: DPO-Tuned for Cyber-Defense Reasoning

This model, kinetigor-dpo-cybersec, is a 0.5 billion parameter language model fine-tuned from Mistral-7B-Instruct-v0.2 using Direct Preference Optimization (DPO). Its core purpose is to enhance cyber-defense and security analyst reasoning capabilities, particularly within a 32768 token context window.

Key Capabilities

  • Specialized Training Data: Trained on a unique, consolidated dataset (nyxspecter4/kin-cyber-all-cybersec-dpo) comprising 443 unique DPO triples. This dataset integrates:
    • MITRE ATT&CK tactics, techniques, mitigations, and groups
    • OWASP Top-10 (web) and OWASP API Top-10 (2023)
    • GraphQL attack surface, NIST CSF, SANS Top-20, CIS Controls v8
    • Sigma rules, CVE, CISA KEV, and Threat-Intel categories.
  • DPO-Tuned for Accuracy: The DPO approach aims for high accuracy in specific cybersecurity reasoning tasks, with a reported DPO-pair-accuracy of 0.85.
  • Competitive Positioning: It carves a niche by focusing on DPO + MITRE + CTI-Bench accuracy, differentiating itself from broader cybersecurity LLMs by its specialized dataset and tuning approach.

Intended Use Cases

  • MITRE ATT&CK TTP Identification: Useful for CTI-ATE (Attack Technique Enumeration).
  • Framework Alignment: Assists in aligning security analysis with OWASP, NIST, and SANS frameworks.
  • Detection Rule Ideation: Supports the generation of ideas for security detection rules.
  • Threat Model Drafting: Aids in the initial drafting of threat models.
  • CTI MCQ Scoring: Can be used for scoring CTI-MCQ (Cyber Threat Intelligence Multiple Choice Questions).

Limitations

Based on Mistral-7B, DPO can sometimes lead to confident but ungrounded claims. It is recommended to pair this model with retrieval-augmented generation (RAG) in production environments. Benchmarking against CTI-Bench is pending.

Finetunes
1 models