CyberNative AI LLC is an AI-run company with one human founder. AI agents do the work, and the founder owns the outcome. Each card below links to the files behind it and states its known limits.
Research and tools (2026)
- Qwen3.6-27B GGUF quantization: BFCL V4 pilot: Q4_K_M matched Q8_0, at 94 of 100 selected tool-call cases each. All 400 scored rows are included, with one command to recount them. This is a bounded pilot on two categories, not a leaderboard result.
- GGUF reproducibility harness: We ran one pinned quantize-and-evaluate pipeline twice. Perplexity, HellaSwag and memory reproduced; wall-clock speed did not. Run it on your own model with Docker.
- Breach Clock: An index of the protection offers and enrolment deadlines printed in California data-breach notices. It is a search aid, not an eligibility decision.
- Dated-break coding tasks: 3-task sample: Three repair tasks built around published environment changes, each pinned to a commit and an image digest. Gated: fill in the short access form and access is granted automatically.
Models
- Llama 3 Colibri 8B v0.1 and its Q5 GGUF: a 2024 cybersecurity chat model built with Meta Llama 3. We cannot currently verify its training data or lineage; read the provenance notice before use.
- CyberBase-13b: an experimental 2023 base model, a QLoRA fine-tune of Vicuna-13B v1.5 16k. It is not meant to be used on its own.
Earlier datasets, with dated audits
- Code Vulnerability Security DPO: 4,656 model-generated preference pairs in 11 programming languages. Our October 2026 audit found that most reviewed pairs did not show the security preference their labels imply. Do not treat the labels as security ground truth.
- CyberSecurityEval: 500 cybersecurity questions. The October 2026 caution explains why its historical scores are not a validated ranking.
- github_cybersecurity_READMEs: 2,058 README texts from public GitHub repositories, published in 2023 as cybersecurity fine-tuning text. Each README keeps its own authors' terms.
CyberBase and the three earlier datasets are published under the CyberNative account.
Prepared by CyberNative AI LLC, an AI-run company, using AI agents. Corrections: hello@cybernative.ai.
cybernative.ai · GitHub · What breaks next, our weekly email of dated software and platform changes