This repository is publicly accessible, but you have to accept the conditions to access its files and content. By agreeing you accept to share your contact information (email and username) with the repository authors. This dataset contains cybersecurity offensive techniques and exploit-related content. Access is granted only for legitimate research and educational purposes.

You agree to use this dataset solely for lawful, defensive security research and education. You will NOT use this data for unauthorized access, malicious attacks, or any illegal activity. You acknowledge that the dataset contains offensive security content collected from public sources for the purpose of training AI systems to better understand and defend against cyber threats. You must cite this dataset in any publication or derivative work.

CyberSecurity-1M

The largest open cybersecurity continued pre-training dataset

Why CyberSecurity-1M?

Existing LLM pre-training datasets (RedPajama, FineWeb, The Stack) contain minimal cybersecurity content — typically <0.1%. This creates a knowledge gap where models lack foundational understanding of:

How exploits work and why they succeed
Step-by-step attack reasoning chains (not just tool output)
Vulnerability classification and root cause analysis
Defensive countermeasures and secure coding patterns

CyberSecurity-1M fills this gap with 1,039,474 raw records from 126 authoritative sources, processed through a 5-stage "Data Darwinism" pipeline that classifies, deduplicates, and enriches content for maximum training utility.

Quick Start

from datasets import load_dataset

# Load L4 refined data (recommended for training)
ds = load_dataset("morinoppp/CyberSecurity-1M", "l4_ctftime", split="train")

# Load raw data from a specific source
ds = load_dataset("morinoppp/CyberSecurity-1M", "raw_hacktricks_cloud", split="train")

# Load all L4 processed records (combined)
ds = load_dataset("morinoppp/CyberSecurity-1M", "processed", split="train")

Dataset Structure

Tier	Path	Records	Description
Raw	`raw/{source}/data.jsonl`	1,039,474	Original crawled data from 126 sources
L3 Classified	`processed/l3/{source}.jsonl`	364,883	Filtered, classified, educational value scored
L4 Refined	`processed/l4/{source}.jsonl`	169,997	Noise removed, structure enhanced
L5 Cognitive	`processed/l5/{source}.jsonl`	285	Reasoning chains completed (high-value subset)
Combined	`processed/train.jsonl`	169,997	All L4 processed records in one file

Pipeline Overview

The 5-stage "Data Darwinism" pipeline progressively refines data quality:

Raw (1039K+) → L1 Normalize → L2 Filter → L3 Classify → L4 Refine → L5 Cognitive
                                                ↓
                                          ┌─────┴──────┐
                                          │ edu 1-2:   │
                                          │ L4 only    │
                                          └─────┬──────┘
                                          ┌─────┴──────┐
                                          │ edu 3-5:   │
                                          │ L4 + L5    │
                                          └────────────┘

Stage	Name	Input → Output	What It Does
L1	Normalize	Raw → Cleaned	Format unification, encoding fixes, field extraction
L2	Filter	Cleaned → Filtered	Garbled text removal, short content filter, MinHash dedup
L3	Classify	Filtered → Classified	LLM educational value scoring (1-5), category assignment, routing
L4	Refine	Classified → Refined	Noise removal, structure enhancement, code formatting
L5	Cognitive	High-value → Enriched	LLM reasoning chain completion, terminology explanation

Source Categories

Category	Sources	Records	Size
CTF & Challenges	6	76,986	1905 MB
Exploits & Vulnerabilities	10	826,804	2139 MB
Offensive Techniques	10	3,496	40 MB
CTI & Threat Intel	7	14,661	1256 MB
Cloud & Container	4	124	1 MB
AI/LLM Security	2	199	2 MB
Detection & Response	5	21,834	148 MB
Security Research	12	18,846	184 MB
Vendor Security Blogs	12	2,017	316 MB
Chinese Security Community	11	55,835	851 MB
Tools & References	8	4,251	112 MB
Other	39	14,421	406 MB

Top Raw Sources

Source	Records	Size
nvd_cve	347,708	1234 MB
osv	259,859	402 MB
trickest_cve	157,961	258 MB
exploitdb	46,553	204 MB
expku	45,522	815 MB
ctftime	38,830	1481 MB
ctf_search	35,803	320 MB
github_security_md	17,440	133 MB
nuclei_templates	13,068	81 MB
vcdb	10,038	45 MB
poc_in_github	8,776	13 MB
pentester_land	6,402	5 MB
xianzhi	5,820	3 MB
seclists	5,259	49 MB
xz_aliyun	4,000	4 MB

Top L4 Refined Sources

Source	Records	Size
go_exploitdb	53,313	139 MB
expku	31,590	520 MB
ctf_search	28,642	504 MB
gh_advisories	24,292	118 MB
ctftime	18,548	567 MB
infocon	2,018	71 MB
pstips	1,070	35 MB
pentest_partners	961	19 MB
exploitdb	956	1 MB
hacktricks	821	16 MB
cisa_kev	798	2 MB
blackhills	667	14 MB
pentesterlab	620	6 MB
ctf_wiki	594	11 MB
oxdf	573	74 MB
atomic_red_team	330	6 MB
awesome_list	330	23 MB
security_rss	292	93 MB
apt_notes	285	20 MB
gtfobins	285	1 MB
kali_tools_zh	262	4 MB
ired_team	236	3 MB
claroty	228	10 MB
websec	219	1 MB
hacker_recipes	209	2 MB
lolbas	209	1 MB
netresec	197	3 MB
quick_reference	194	19 MB
skullsecurity	161	6 MB
awesome_web_security	144	4 MB
loldrivers	123	0 MB
websec_notes	116	2 MB
gh_security_lab	107	1 MB
eset_welivesecurity	99	1 MB
h4cker	91	4 MB
cloudgoat	66	0 MB
local_books	54	75 MB
bishop_fox	42	1 MB
emergency_response	41	1 MB
liveoverflow	38	1 MB
sql_injection_wiki	21	0 MB
burpsuite_guide	20	0 MB
elastic_security	20	1 MB
freebuf	19	0 MB
antiy_ti	13	1 MB
cisco_talos	13	1 MB
apple_security	10	0 MB
cybereason	10	0 MB
mitre_attack	10	7 MB
fortinet_fortiguard	9	1 MB
binary_ninja	8	1 MB
eclypsium	7	0 MB
cert360	5	0 MB
anquanke_deep	3	0 MB
checkpoint_research	2	0 MB
akamai_blog	1	0 MB
crowdstrike_blog	1	0 MB
cxsecurity	1	0 MB
fsecure	1	0 MB
offsec_msfu	1	0 MB
sqlmap_wiki	1	0 MB

Data Fields

Raw Records

Field	Type	Description
`id`	string	Unique record identifier
`title`	string	Record title
`source`	string	Data source identifier
`url`	string	Original URL
`markdown`	string	Main content in Markdown
`raw_html`	string	Original HTML (optional)
`description`	string	Short description (optional)
`category`	string	Source-specific category
`tags`	list[str]	Topic tags
`author`	string	Author (optional)
`date`	string	Publication date (optional)
`cve`	string	CVE identifier (optional)
`extra`	dict	Source-specific metadata
`scraped_at`	string	ISO 8601 collection timestamp

L4 Refined Records

All raw fields plus:

Field	Type	Description
`markdown_l4`	string	L4 refined content (cleaned, restructured)
`l3_category`	int	L3 category code
`l3_category_name`	string	L3 category name
`l3_content_types`	list[str]	Content type tags
`l3_edu_value`	int	Educational value score (1-5)
`l3_noise_types`	list[str]	Noise types identified
`l3_route`	string	Pipeline route: "l4_only" or "l4_l5"
`l4_issues`	list[str]	Issues found and fixed in L4

Access Requirements

This is a gated dataset. To access the data, you must:

Provide your Name, Institution, Country, and Email
Describe your Intended Use of the dataset
Agree to the terms of use

Terms of Use

By accessing this dataset, you agree to:

Use the data solely for research and educational purposes
Not use the data for malicious activities, unauthorized access, or any illegal purpose
Respect the original copyright of individual content sources
Cite this dataset when using it in publications
Not redistribute the dataset without permission

Citation

@dataset{cybersecurity_1m_2026,
  title={CyberSecurity-1M: A Million-Scale Cybersecurity Continued Pre-Training Dataset},
  author={morinoppp},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/morinoppp/CyberSecurity-1M}
}

License

Apache License 2.0. Individual content retains its original copyright. This dataset is provided for research and educational purposes only.

Downloads last month: 21

Total file size:

20.9 GB