README.md

---
license: other
task_categories:
- text-retrieval
- text-generation
- question-answering
- summarization
- feature-extraction
language:
- en
tags:
- security
- red-team
- redteam
- penetration-testing
- pentesting
- web-security
- cloud-security
- active-directory
- kubernetes
- dfir
- cheatsheets
- wiki
- markdown
- jsonl
- cybersecurity
- cyber-security
- dataset
- retrieval
pretty_name: HackTricks SlimPajama Corpus
size_categories:
- n<1K
dataset_info:
  config_name: hacking_tricks
  features:
  - name: text
    dtype: string
  - name: meta
    struct:
    - name: source
      dtype: string
    - name: source_repo
      dtype: string
    - name: repo_branch
      dtype: string
    - name: repo_commit
      dtype: string
    - name: source_path
      dtype: string
    - name: source_url_github
      dtype: string
    - name: source_url_site
      dtype: string
    - name: title
      dtype: string
    - name: section_titles
      list: string
    - name: license
      dtype: string
    - name: n_chars
      dtype: int64
    - name: n_words
      dtype: int64
    - name: sha256
      dtype: string
    - name: crawl_ts
      dtype: string
    - name: lang
      dtype: string
  splits:
  - name: train
    num_bytes: 9726009
    num_examples: 888
  download_size: 4839940
  dataset_size: 9726009
configs:
- config_name: hacking_tricks
  data_files:
  - split: train
    path: hacking_tricks/train-*
---

# HackTricks SlimPajama Corpus (Parquet)

A high‑quality, SlimPajama‑style corpus distilled from the **HackTricks** knowledge base (core wiki and, optionally, HackTricks Cloud). Each record corresponds to a cleaned Markdown article with headings preserved, code fences kept, and rich provenance metadata for reproducible research and downstream retrieval.

**Token count:** ~2.75M tokens.


> ⚠️ **Ethical use only.** This corpus is intended for lawful research, education, and defensive security. Do **not** use it to violate terms of service or applicable laws.

---

## What’s in this release (Parquet)

* **Primary delivery = Parquet** shards under `data/hacking_tricks/.../train-*.parquet` (Hub‑managed) for fast streaming with 🤗 `datasets`.
* **Raw JSONL** retained at `raw/hacking-tricks.jsonl[.gz|.zst]` for transparency and reproducibility.
* **Named config:** `hacking_tricks` (single config; **no** `all` config).

---

## Highlights

* **Clean, article‑level records** suitable for pretraining/continued‑pretraining, RAG, and eval.
* **Code blocks preserved** (fenced code kept as‑is) while removing front‑matter/boilerplate.
* **Provenance**: repo, commit, path, GitHub blob URL, and a best‑effort canonical site URL per article.
* **Lightweight schema**: `{"text": ..., "meta": {...}}` with stable keys for training and retrieval.

---

## Dataset layout

```
/                           # dataset root (this README)
  raw/
    hacking-tricks.jsonl    # original export (optionally .gz/.zst)
  data/
    hacking_tricks/
      default/1.0.0/train/
        train-00000-of-XXXX.parquet
        train-00001-of-XXXX.parquet
        ...
```

> Depending on future pushes, you may also publish separate configs (`hacktricks_core`, `hacktricks_cloud`). This card documents the current **single** `hacking_tricks` config.

---

## Schema

Each row is a UTF‑8 JSON object with the following structure (identical for Parquet features):

```json
{
  "text": "<cleaned article markdown with code fences>",
  "meta": {
    "source": "HackTricks (git)",
    "source_repo": "https://github.com/HackTricks-wiki/hacktricks",
    "repo_branch": "master",
    "repo_commit": "<git-sha>",
    "source_path": "path/in/repo/file.md",
    "source_url_github": "https://github.com/.../blob/<sha or branch>/path.md",
    "source_url_site": "https://book.hacktricks.wiki/<page>",
    "title": "<article title>",
    "section_titles": ["H1/H2/H3…"],
    "n_chars": 12345,
    "n_words": 2345,
    "sha256": "<content-hash>",
    "crawl_ts": "2025-11-04T00:00:00+00:00",
    "lang": "en",
    "license": "CC BY-NC-SA 4.0 (per upstream repo)"
  }
}
```

**Field notes**

* `text` keeps fenced code blocks and normalized headings. Front‑matter is removed.
* `source_url_site` is a best‑effort canonical link to the rendered wiki page; `source_url_github` always points to the exact file/commit.
* `license` mirrors the upstream license hint for clarity.

---

## Quickstart (🤗 Datasets)

### Load the Parquet config (recommended)

```python
from datasets import load_dataset

REPO = "tandevllc/hacking-tricks"
ds = load_dataset(REPO, name="hacking_tricks", split="train")
print(len(ds), ds[0]["meta"]["title"])  # sample
```

### Direct JSONL loading from `raw/` (explicit)

```python
from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="raw/hacking-tricks.jsonl",
    repo_id="tandevllc/hacking-tricks",
    split="train",
)
```

### Typical filters

````python
# Long English articles only
long_en = ds.filter(lambda r: r["meta"].get("lang") == "en" and r["meta"].get("n_words", 0) > 800)

# Articles with code (simple heuristic)
with_code = ds.filter(lambda r: "```" in (r["text"] or ""))

# Slice by section heading keyword
ad_ops = ds.filter(lambda r: any("Active Directory" in h for h in r["meta"].get("section_titles", [])))
````

### RAG‑ready retrieval columns

* `meta.source_url_github` — stable pointer to the exact file.
* `meta.source_url_site` — user‑friendly URL for UI previews.
* `meta.sha256` — robust dedup key across updates.

---

## Intended uses

* **Pretraining / Continued pretraining** for security‑aware models.
* **RAG / QA** over a curated set of red‑team/blue‑team encyclopedic pages.
* **Summarization & explanation** of workflows, checklists, and attack surfaces.
* **Evaluation**: security knowledge, long‑form reasoning on procedures, and code‑aware contexts.

> This corpus is **didactic/encyclopedic**. It does **not** replace hands‑on lab data, exploit PoCs, or live telemetry.

---

## Cleaning & quality

* YAML front‑matter and boilerplate removed; headings normalized; excessive blank lines collapsed.
* Code fences retained verbatim for downstream code‑aware models.
* Article‑level minimum length applied (very short files excluded).
* Stable dedup via `meta.sha256` and exact `source_path` at a specific commit.

---

## Limitations & caveats

* **Upstream drift**: HackTricks evolves; future updates may add/rename pages.
* **License constraints**: Non‑commercial, ShareAlike terms apply; see License below.
* **Context gaps**: Some pages assume prior knowledge or external links.
* **No execution traces**: Articles describe tactics; they are not runtime logs or CVE feeds.

---

## License

* **Underlying content**: © respective HackTricks authors, typically **CC BY‑NC‑SA 4.0** as stated in the upstream repositories/site.
* **License:** "TanDev Proprietary License — All Rights Reserved"

When you use this dataset, you **must**:

1. Provide **attribution** to HackTricks and the specific article(s) used.
2. Use it **non‑commercially** unless you have additional permission from rights holders.
3. Distribute derivatives under the **same license** (ShareAlike).

If you are a rights holder and would like a page revised or removed, please open an issue on the dataset page or email the maintainer.

---

## Citation

If this dataset helps your research, please cite it alongside HackTricks:

```bibtex
@dataset{tandevllc_2025_hacktricks_slimpajama,
  author = {Gupta, Smridh},
  title  = {HackTricks SlimPajama Corpus},
  year   = {2025},
  url    = {https://huggingface.co/datasets/tandevllc/hacking-tricks}
}
```

You should also attribute individual HackTricks pages used (e.g., in an acknowledgments section) to satisfy CC BY‑NC‑SA 4.0.

---

## Maintainer

**Smridh Gupta** — [smridh@tandev.us](mailto:smridh@tandev.us)