HackerSignal

A large-scale, multi-source dataset linking hacker community discourse, exploit databases, vulnerability advisories, and fix commits through a shared CVE identifier space.

Overview

Statistic	Value
Documents	7,447,646 (exact-deduplicated)
Sources	64 public forum/source identifiers
Source layers	8
Temporal span	1988--2026
CVE-linked rows	360,004
Benchmark tasks	3

Quick Start

from datasets import load_dataset

# Load the 10K stratified sample (24 MB)
sample = load_dataset("DatasetSubmission/HackerSignal", "sample")

# Load a benchmark task
task1 = load_dataset("DatasetSubmission/HackerSignal", "task1_cve_linkage")

# Load the full corpus (12 GB)
full = load_dataset("DatasetSubmission/HackerSignal", "default")

Source Layers

Layer	Description	Rows
hacker_community	Forum posts from hacker/security communities	6,974,128
exploit_archive	Published exploit code and advisories	197,393
vulnerability_reference	NVD CVE descriptions and structured advisories	150,916
exploit_qa_reference	Security Q&A and tutorial content	68,791
advisory_reference	Vendor and third-party security advisories	30,361
bug_bounty_disclosure	Public bug bounty reports	13,720
fix_commit_reference	Vulnerability fix commit messages	11,104
exploitation_reference	Active exploitation indicators (CISA KEV)	1,233

Benchmark Tasks

Task 1: CVE Linkage Retrieval (CVE-R)

Cross-source temporally OOD entity grounding: given exploit/advisory evidence text, retrieve the correct NVD CVE entry from a corpus of 340K descriptions. Quality-controlled to remove unretrievable rows.

Split	Rows
Train	40,698
Val	2,551
Test	550
Corpus	340,536

Task 2: Exploit Type Classification (ETC)

8-class temporal OOD classification of exploit posts by type: injection, XSS, memory corruption, DoS, file inclusion, authentication/access bypass, RCE, and information disclosure. Sourced from ExploitDB and HackerOne public reports; split by publication year so test posts postdate all training examples.

Split	Rows
Train	64,413
Val	4,735
Test	1,735

Task 3: Temporal Generalization (TG)

Identical retrieval formulation to Task 1 but with strict CVE-disjoint constraint: C_train ∩ C_test = ∅. Split by CVE publication year (train: <2022, test: 2024+) to test generalization to wholly unseen vulnerabilities.

Split	Rows
Train	41,630
Val	1,776
Test	393
Corpus	340,536

Schema

Each record in the main corpus contains:

Field	Type	Description
unified_id	string	SHA-256-derived release identifier
source_dataset	string	Input source file or dataset name
source_layer	string	Normalized layer (see above)
text	string	UTF-8 post/advisory text (max 8000 chars)
timestamp	string	ISO 8601 publication datetime (UTC)
forum_id	string	Source identifier
author_hash	string	SHA-256 pseudonymized author
release_mode	string	Governance mode (see below)

Release Governance

Records follow source-specific release modes:

• redistributable_text (5 sources): Full text included under public-domain or CC licenses.
• research_text_with_terms (9 sources): Text included with source attribution; no commercial reuse.
• metadata_or_pointer_only (13 sources): Only metadata, timestamps, CVE refs, text hashes, and lengths; raw text withheld.

See docs/release_governance.md for the full source matrix, takedown policy, and responsible-use restrictions.

Responsible Use

This dataset is intended for defensive cybersecurity research only. Prohibited uses include:

• Training systems for automated exploit code generation or malware
• De-anonymizing forum participants outside approved research protocols
• Operational blocking or law-enforcement decisions based solely on model outputs
• Republishing text from metadata_or_pointer_only sources

Known Issues

• HuggingFace dataset viewer row count: The viewer may display a partial count (~1M) for the default/full split because HF's streaming viewer cannot fully index files over ~10 GB. The actual row count is 7,447,646 as validated against the SHA-256 manifest.
• Content-scanner warnings: HF's automated scanner flags several files (including the sample JSONL) as "Unsafe" due to exploit and vulnerability text. This is expected for a cybersecurity research dataset and does not indicate malicious content. See the Responsible Use section below.

Limitations

• English-language bias; non-English communities may be underrepresented
• Historical forum imports skew toward 2015--2017
• Task labels are weak (CVE-linkage-derived), not human-annotated
• ~8% near-duplicate content remains after exact deduplication
• 1.57M rows under 8 tokens (useful for interaction analysis, not classification)

Citation

@misc{hackersignal2026,
  title={{HackerSignal}: A Temporally Structured, Multi-Source Benchmark for Cybersecurity Threat Intelligence},
  author={Anonymous},
  year={2026},
  howpublished={arXiv preprint}
}

License

Code and metadata: CC BY 4.0. Source text retains source-specific terms documented in the governance addendum and release manifest.