Dataset Card for RedSage-CFW

RedSage: A Cybersecurity Generalist LLM" (ICLR 2026).
Authors: Naufal Suryanto¹, Muzammal Naseer^1†, Pengfei Li¹, Syed Talal Wasim², Jinhui Yi², Juergen Gall², Paolo Ceravolo³, Ernesto Damiani^3
1Khalifa University, ²University of Bonn, ³University of Milan
^†Project Lead


🌐 Project Page  |   🤖 Model Collection  |   📊 Benchmark Collection  |   📘 Data Collection

---

Dataset Summary

RedSage-CFW (CyberFineWeb) is a large-scale, cybersecurity dataset designed for the continual pretraining of Large Language Models (LLMs). It consists of approximately 11.7 billion tokens spanning 13 million documents.

The dataset was constructed by filtering the FineWeb corpus (Common Crawl 2013–2024) using a custom ModernBERT-based classifier to identify cybersecurity-relevant content. To prevent catastrophic forgetting of general capabilities during pretraining, the cybersecurity data is mixed with general educational content from FineWeb-Edu.

Supported Tasks

• Continual Pretraining: Designed to adapt general-purpose LLMs (e.g., Qwen, Llama) to the cybersecurity domain.
• Domain Adaptation: Enhances model performance on cybersecurity knowledge, skills, and tool usage

Languages

The dataset primarily consists of English text, derived from Common Crawl sources.

Dataset Structure

Data Instances

The dataset is partitioned into 5 chunks (config names: chunk_1 through chunk_5). Each instance represents a single document (e.g., a web page, article, or forum post).

Data Fields

Based on the provided configuration, the data fields are:

• text (string): The full text content of the document.
• id (string): A unique identifier for the document.
• metadata (struct): Contains detailed attributes about the source and filtering:
• probability (float64): The confidence score from the cybersecurity classifier.
• relevant (bool): A flag indicating if the document passed the relevance filter.
• url (string): The source URL of the document.
• date (timestamp): The crawl or publication date.
• dump (string): The Common Crawl dump identifier (e.g., CC-MAIN-2024-51).
• file_path (string): Path information for the original file.
• language (string): The detected language of the text.
• language_score (float64): Confidence score of the language detection.
• token_count (int64): The number of tokens in the document.
• score, int_score: Additional quality or relevance metrics.

Data Splits

The dataset is segmented into 5 chunks. The paper notes that the final corpus consists of the "latest 5 chunks" from the filtered pipeline to fit training budgets.

• Total Size: ~11.7B tokens.
• Total Documents: ~13M documents.

Dataset Creation

Curation Rationale

Existing cybersecurity solutions often rely on proprietary APIs or lack domain adaptation. RedSage-CFW bridges this gap by providing a transparent, open-source corpus for training local, privacy-preserving cybersecurity assistants.

Source Data

• FineWeb: The base corpus is FineWeb, aggregated from 104 Common Crawl subsets between Summer 2013 and December 2024 (~17.2T tokens).
• FineWeb-Edu: Used for mixing general knowledge to maintain reasoning capabilities.

Data Processing & Filtering

1. Classifier Training: A binary classifier based on ModernBERT-base was trained on the "Cybersecurity Topic Classification" dataset (sourced from Reddit, StackExchange, and arXiv). It achieved 97.3% accuracy on validation.
2. Filtering: This classifier was applied to FineWeb, identifying ~125M cybersecurity-relevant documents (~89.8B tokens).
3. General Knowledge Replay: To avoid catastrophic forgetting, the cybersecurity data was mixed with FineWeb-Edu samples at a 30% replay ratio.
4. Deduplication: Global deduplication was performed using MinHash-LSH (via DataTrove), reducing the corpus size by ~47.9% in tokens.
5. Chunking: The final dataset comprises the latest 5 chronological chunks from the processed data to manage computational costs.

Considerations for Using the Data

Social Impact

The dataset enables the development of open-source cybersecurity assistants, potentially helping to bridge the global skills shortage in the field.

Discussion of Biases and Limitations

• Source Bias: As a web-crawled dataset, it may inherit biases present in Common Crawl and online cybersecurity discussions.
• Dual Use: The dataset may contains offensive security knowledge (e.g., penetration testing techniques). While intended for defense, there is an inherent risk of misuse.

---

Citation

@inproceedings{suryanto2026redsage,
  title={RedSage: A Cybersecurity Generalist {LLM}},
  author={Naufal Suryanto and Muzammal Naseer and Pengfei Li and Syed Talal Wasim and Jinhui Yi and Juergen Gall and Paolo Ceravolo and Ernesto Damiani},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=W4FAenIrQ2}
}