license: other
task_categories:
- text-generation
language: - en
tags: - security
- red-team
- prompt-injection
- distillation
- cybersecurity
size_categories: - n<1K
RedSec Distillation SFT
A clean, distilled supervised fine-tuning dataset for security-focused language models. Built for authorized red-team, penetration-testing, and LLM-security research, and used to train RedSec-7B and RedSec-14B on Hugging Face.
Why this dataset is different
Most security instruction datasets pair prompts with raw, templated payloads. Fine-tuning on those teaches a model the format but fills it with noise. This one flips that: every answer was generated by a capable base model and quality-filtered, so each training target is both correct and consistently voiced. Garbage in, garbage out, so only the good answers were kept.
At a glance
| Examples | 794 (732 train / 31 validation / 31 test) |
| Format | chat messages (user + assistant turns) |
| Domains | prompt injection, web payloads (SQLi / XSS / SSRF / XXE / ...), CVE & vulnerability Q&A, secure-code fixes |
| Method | distillation + curated OWASP LLM01 gold |
| Models trained | RedSec-7B, RedSec-14B |
| License | responsible-use (details below) |
Format
Each line is one JSON object:
{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
Quick start
from datasets import load_dataset
ds = load_dataset("json", data_files="train.jsonl")["train"]
print(ds[0]["messages"])
How it was built
- A teacher model (WhiteRabbitNeo-2.5 / Qwen2.5-Coder-7B) answered a diverse pool of authorized security-testing prompts in a fixed RedSec voice.
- Every answer was filtered: refusals, near-empty outputs, and degenerate or looping text were dropped (about 99% passed).
- Curated gold (correct, universal OWASP LLM01 prompt-injection techniques) was added and upweighted to sharpen nuanced categories.
Models trained from this dataset
| Model | Notes |
|---|---|
| redsec-7b-distill | 24-prompt eval: 0 refusals, 0 looping, 24/24 on-voice |
| redsec-14b-distill | the 14B step up for nuanced attacks |
Intended use and ethics
For authorized security testing, red-teaming, and defensive research (for example, training prompt-injection detectors or robustness benchmarks). This dataset contains live attack strings. Not for unauthorized access or harm. Use only on systems you own or are explicitly authorized to assess. Content is synthetically generated and may contain inaccuracies, so verify any payload before use.
Provenance and license
Answers are outputs of WhiteRabbitNeo-2.5 (base Qwen2.5-Coder-7B, Apache-2.0, with a responsible-use extension); prompts were sampled from public, permissively licensed security-testing instructions. Released under a responsible-use license: permissive for authorized security research and defensive work, no unauthorized or harmful use.