RedTeam/README.md

metadata

license: mit
task_categories:
  - text-generation
  - question-answering
language:
  - en
size_categories:
  - 10K<n<100K
pretty_name: RedTeam-Premium
tags:
  - red-team
  - offensive-security
  - penetration-testing
  - cybersecurity
  - vulnerability-research
  - ctf
  - security
  - sft
  - fine-tuning
  - distillation
  - alpaca
  - knowledge-base
  - mit-license

🗡️ RedTeam-Premium

A deduplicated, quality-filtered, instruction-SFT-ready red-team dataset of 19,033 traces, converted from the raw WNT3D Ultimate Red Team collection into a single clean format. Part of the Premium series — see fable-5-premium, fable-5.1-premium, CyberSec-Reasoning-Premium, Kimi-K3-Premium, and Qwen3.8-Agent-Premium.

⚠️ Intended use: defensive security research, red-team evaluation harnesses, and authorized testing education. Do not use for unauthorized access against systems you do not own or have permission to test.

Priorities: Quality > Ease of Access > Quantity

📊 Dataset Overview

Property	Value
Total Traces	19,033
Train Split	16,178 (85.0%)
Validation Split	1,427 (7.5%)
Test Split	1,428 (7.5%)
Average Quality	0.62 (dedup + oversize cap applied)
Source	WNT3D/Ultimate-Offensive-Red-Team (MIT)
Language	English only
Created	2026-09-20

🔧 What we did (and why this is a different repo than the source)

The source is a valuable but raw collection: 6 overlapping Alpaca-JSONL files with 27,386 exact duplicate rows, generic wrapper instructions ("User query", "Answer the question"), 66% of rows with sub-100-char outputs, and several 6.5MB single-row JSON dumps. This repo is the claimable clean version:

1. Deduplication — SHA-256 on normalized instruction+output; 27,000+ cross-file duplicates collapsed
2. Oversize Cap — Rows with >60KB outputs dropped (SFT-unusable mega-dumps)
3. Instruction Normalization — Generic wrappers folded into real task content
4. Knowledge-Base Conversion — 8 structured JSON knowledge bases (attack playbooks, vulnerability database, tool/exploit references, DEFI exploits, rust security, smart-contract vulns, target analysis, operational framework) converted into instruction→explanation rows
5. PII Scrubbing — Paths, API keys, emails, IPs replaced with [REDACTED_*]
6. Quality Filtering — Short (<200 char) and sub-threshold rows dropped

Source yield: 78,430 raw rows → 19,033 kept (24.3%).

📈 Coverage

Initial access · privilege escalation · lateral movement · persistence · exfiltration · defense evasion · vulnerability database (KEV) · tool usage (ghidra, burpsuite, metasploit, nmap, responder, bloodhound, mimikatz) · DEFI/smart-contract exploits · rust memory-safety · operational decision framework

🎯 Usage

from datasets import load_dataset

ds = load_dataset("saidutta69/RedTeam-Premium", data_files="openai_chat/train.parquet", split="train")

With Axolotl

datasets:
  - path: saidutta69/RedTeam-Premium
    type: alpaca
    data_files: openai_chat/train.parquet
    split: train

🧰 Schema

Both formats (openai_chat/ and agent_traces/) carry identical rows: messages (JSON string, user/assistant), model, source, session_id, quality_scores (JSON string).

🧾 Provenance & License

MIT (inherited from WNT3D/Ultimate-Offensive-Red-Team). Full build stats shipped in build_stats.json. Built through the premium cleaning pipeline from the listed public source. Companion datasets: CyberSec-Reasoning-Premium · Kimi-K3-Premium · Qwen3.8-Agent-Premium.