Fenrir v2.1: 100K Defensive Cybersecurity Chat Triples for Safe LLM Fine‑Tuning
AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1 ↗
The Fenrir v2.1 dataset, authored by Alican Kiraz, offers 99,870 high‑quality *system / user / assistant* chat triples designed for instruction‑tuning of defensive‑security language models. All records are in English, stored as JSONL, and follow a strict schema that includes a system prompt describing a seasoned cyber‑defense AI, a user query, and an assistant response that provides mitigation guidance or a safe refusal. The dataset is Apache‑2.0 licensed, making it commercial‑friendly.
Coverage spans the major security frameworks—OWASP Top 10, MITRE ATT&CK, NIST CSF, CIS Controls, ASD Essential 8—as well as modern topics such as cloud IAM, DevSecOps pipelines, OAuth2/OIDC/SAML authentication, TLS/cryptography hygiene, and AI‑security interplay. Each assistant answer is crafted at a senior security engineer level and includes built‑in refusal patterns to prevent dual‑use or exploit‑building requests. Quality controls include deduplication, PII scrubbing, hallucination scans, and adversarial refusal tests.
Because the data is organized as chat triples, it can be directly fed into Hugging Face’s `datasets` library (or pandas/polars) for supervised fine‑tuning of LLMs on defensive‑security tasks. The dataset’s alignment‑safe design also makes it suitable for evaluating or benchmarking safety mechanisms in security‑focused conversational agents. With a single train split and a modest size (under 100 K rows), Fenrir v2.1 is ready for rapid experimentation and production deployment.
Project Ideas
- Fine‑tune a conversational LLM to act as a cyber‑defense assistant that answers security mitigation questions while safely refusing malicious requests.
- Create a security‑policy checklist generator that takes a user‑provided scenario and returns step‑by‑step hardening actions mapped to OWASP, NIST, or CIS controls.
- Build a retrieval‑augmented QA system that indexes Fenrir’s triples and provides context‑rich answers for cloud IAM or DevSecOps best practices.
- Evaluate the robustness of existing security chatbots by testing them on Fenrir’s refusal‑pattern prompts and measuring alignment compliance.
- Develop an interactive training tool for junior security engineers that presents user prompts from the dataset and asks learners to draft their own mitigation responses.