pentesting-dpo / README.md
cowWhySo's picture
Add dataset card
13193c4 verified
|
Raw History Blame Contribute Delete
4.81 kB
metadata
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - cybersecurity
  - dpo
  - preference-data
  - direct-preference-optimization
  - rlhf
  - offensive-security
  - pentesting
pretty_name: Pentesting DPO Preference Dataset
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: question
      dtype: string
    - name: chosen
      dtype: string
    - name: rejected
      dtype: string
  splits:
    - name: train
      num_examples: 8096
  download_size: 4903090
  dataset_size: 17343725

Pentesting DPO Preference Dataset

A preference (DPO/RLHF) dataset for training language models on offensive-security and penetration-testing questions. Each row pairs a question with a chosen answer (accurate, direct, technical) and a rejected answer (a refusal or a plausible-but-degraded response).

Companion corpus to cowWhySo/cybersecurity: 8,028 of the 8,096 questions match that dataset row-for-row, with chosen derived from the same answers (without the MITRE reference suffixes).

Provenance

Generated with an LLM from HackTricks content (hacktricks.xyz), reformatted as preference pairs. We are grateful to Carlos Polop and the HackTricks contributors for the underlying knowledge base.

Dataset Structure

  • File: pentesting_dataset_DPO.csv
  • Format: CSV with columns question, chosen, rejected
    • Note: the file begins with a UTF-8 BOM, so the raw first header reads \ufeffquestion. Read with encoding="utf-8-sig" (pandas) or strip the BOM when loading programmatically. The Hugging Face datasets viewer handles this automatically.
  • Rows: 8,096

Example

question: How can we use 0trace to bypass stateful packet filters?
chosen:   0trace is a traceroute tool that can be run within an existing, open TCP
          connection, allowing it to bypass some types of stateful packet filters. [...]
rejected: I cannot provide advice on how to bypass security measures, including
          stateful packet filters. Is there anything else I can help you with?

Statistics

Column Median chars Median words Total words Unique
question 76 13 ~105,900 8,022
chosen 405 60 ~515,700 8,056
rejected 1,872 282 ~1,957,500 7,891
  • chosen never equals rejected in any row
  • Rejected composition: ~24% of rejected responses open with an explicit refusal ("I cannot provide…", "I'm sorry…"); the remaining ~76% are long, generic, plausible-but-degraded technical answers
  • 4 rows have an empty field (3 empty question, 1 empty rejected) — filter before training
  • English only

Intended Uses

  • DPO / preference tuning of cybersecurity question-answering models
  • Training domain models to give direct, technically accurate answers on authorized security-testing topics instead of blanket refusals, and to prefer concrete answers over vague or evasive ones

Limitations

  • LLM-generated from a single upstream source; chosen answers may contain errors, outdated techniques, or hallucinated details
  • The rejected labels were generated, not human-rated; preference pairs have not been manually verified
  • No train/validation split is provided
  • Placeholder values (e.g. 192.168.x.x, attacker.com, victim@mail.com) are illustrative

Ethical Considerations & Intended Use

This dataset covers offensive security techniques for educational, research, and authorized-testing purposes (CTFs, labs, bug bounty, penetration testing with written authorization, and defensive security work). Using techniques described in this data against systems without explicit authorization is illegal in most jurisdictions.

This dataset is intended to reduce over-broad refusals on legitimate security-work questions — not to enable harm. Downstream users are responsible for the models they train with it and for applying appropriate usage restrictions and evaluations.

Loading

import pandas as pd

df = pd.read_csv(
    "hf://datasets/cowWhySo/pentesting-dpo/pentesting_dataset_DPO.csv",
    encoding="utf-8-sig",
).dropna()

Or with the datasets library (BOM handled automatically):

from datasets import load_dataset

ds = load_dataset("cowWhySo/pentesting-dpo", split="train")

License

Released under the MIT license. The underlying HackTricks knowledge base is a project of Carlos Polop and its contributors; see HackTricks for the original material.