OffSec RedTeam Codes is a curated corpus of code (and some auxiliary text) extracted from popular GitHub repositories related to offensive security / red teaming (pentesting, OSINT, C2, privilege escalation, exploitation, forensics, etc.). It is also the largest open-source dataset of red-team and offensive-security code ever compiled.

⚠️ Ethical use only. This dataset is for research, education, and defensive security testing in lawful environments with proper authorization. Do not use it to harm systems, violate terms of service, or break the law.

What’s new (Nov 2025)

Parquet per‑topic configs are now published for fast loading with datasets.load_dataset(...).
raw/ folder added: original JSON/JSONL/NDJSON sources moved under raw/ (e.g. raw/github_topic_code-<topic>.jsonl).
All Parquet splits include a constant topic column.
Shards are sized at ≈256–512 MB to balance memory, upload speed, and downstream streaming.

Repository layout

/               # root
  README.md     # this card
  raw/          # original source JSON/JSONL/NDJSON (one file per topic)
  <topic>/      # per-topic Parquet config folder produced by HF Datasets
    train-0000X-of-00N.parquet

Schema

All examples share the same columns (Parquet + raw JSON):

content (string) — full file text (notebook outputs stripped; execution_count cleared)
repo_name (string) — owner/name
repo_id (int) — GitHub repo id
path (string) — file path in repo
branch (string) — default branch at fetch time
license (string|null) — SPDX id or GitHub license key
lang (string) — approx language/extension
stargazers (int) — star snapshot
pushed_at (string, ISO‑8601) — last push timestamp
topics_matched (list[string]) — discovery topics that matched this repo
size (int) — character length of content
topic (string) — added during build, equals the topic/config name

Load the Parquet configs (recommended)

Single topic

from datasets import load_dataset

# Loads Parquet for the selected topic (config name)
ds = load_dataset(
    "tandevllc/offsec_redteam_codes",
    name="pentesting",    # pick any topic listed above
    split="train",
)
print(len(ds), ds.column_names)

Multiple topics (interleaved)

from datasets import load_dataset, interleave_datasets

names = ["osint", "pentesting", "reverse-engineering"]
parts = [load_dataset("tandevllc/offsec_redteam_codes", name=n, split="train") for n in names]

# Optionally balance by size:
probs = [1/len(parts)]*len(parts)
ds = interleave_datasets(parts, probabilities=probs, seed=42)

Filter by language or license

py = ds.filter(lambda r: r.get("lang") == "py")
permissive = ds.filter(lambda r: (r.get("license") or "").lower() in {"mit","apache-2.0","bsd-3-clause"})

Load the raw JSON/JSONL sources (optional)

If you prefer the original line-delimited JSON:

from datasets import load_dataset

raw = load_dataset(
    "json",
    data_files="raw/github_topic_code-pentesting.jsonl",
    split="train",
    repo_id="tandevllc/offsec_redteam_codes",
)

Content selection and filtering

File types: code-like extensions only (py, pyw, pyi, js, mjs, cjs, ts, tsx, jsx, java, kt, kts, groovy, scala, c, h, cpp, cc, cxx, hpp, hh, hxx, ino, rs, go, rb, php, phtml, cs, fs, vb, swift, m, mm, sh, bash, zsh, ksh, fish, ps1, psd1, psm1, sql, pl, pm, r, jl, lua, dart, zig, nim, yaml, yml, toml, ini, cfg, conf, env, properties, gradle, cmake, make, mk, mkfile, json, ipynb).
Excluded paths: build outputs, vendored folders, caches, VCS dirs, virtualenvs
Binaries/archives/media: filtered by extension + text heuristics
Large files: default cap ≈2 MB (configurable)
Minified blobs: excluded by average line-length heuristics
Secrets: basic regex heuristics (keys, private certs) excluded where detected

Intended uses

Pretraining / continued pretraining of code LMs in the security domain
RAG / retrieval with metadata filters (language, topics, license, stars)
Code search / similarity tasks
Curriculum sampling by lang, stargazers, topic, or recency (pushed_at)

This dataset is not an instruction set for unauthorized intrusion. Use defensively and ethically.

Limitations & caveats

Licenses vary per file. The compilation license does not replace upstream licenses.
Heuristics can omit useful files or miss some undesirable ones.
Time-varying metadata (stars, topics, default branch) reflect the crawl moment.

License & Access

Upstream Licenses: Each record includes source_repo and source_license. Rights in the original repositories remain with their owners. This compilation license does not limit rights you already have in the upstream sources.

If you are a repository author and would like your content removed, please open an issue or contact the maintainer.

Citation

@dataset{tandevllc_2025_offsec_redteam_codes,
  author = {Gupta, Smridh},
  title  = {OffSec RedTeam Codes},
  year   = {2025},
  url    = {https://huggingface.co/datasets/tandevllc/offsec_redteam_codes}
}

Maintainer

Smridh Gupta — smridh@tandev.us

Downloads last month: 163