RISys-Lab/RedSage-CFW-Full-Dedup

Dataset Card for RedSage-CFW Full (Dedup)

"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 Full (Dedup) is the globally deduplicated cybersecurity web corpus extracted from FineWeb across all 103 Common Crawl dumps from CC-MAIN-2013-20 through CC-MAIN-2024-51.

Starting from the raw ~122.15M classified documents (~89.8B tokens), this corpus was processed using global MinHash-LSH deduplication via Hugging Face's datatrove library, removing ~47.9% redundant tokens across dumps. The result is packaged as 40 sharded jsonl.gz files totaling ~88.8 GB of compressed data, containing clean, deduplicated domain-specific text spanning over a decade of public cybersecurity discourse.

Note on CFW Versions:

• RedSage-CFW Full (Non-Dedup): Complete raw classifier outputs across 103 dumps (122.15M rows, ~267 GB Parquet). No deduplication applied.
• RedSage-CFW Full (Dedup) [This Dataset]: Complete full-span filtered corpus after global MinHash-LSH deduplication (40 compressed jsonl.gz shards, ~88.8 GB compressed). No general replay mixed.
• RedSage-CFW: The final training-ready corpus combining the latest 5 chronological chunks with MinHash deduplication and 30% FineWeb-Edu general knowledge replay (~13M documents, ~11.7B tokens).

Supported Tasks

• Continual Domain Pretraining: Train or continue pretraining foundational LLMs on clean, non-redundant cybersecurity web text.
• Custom Replay & Mixing Experiments: Mix with custom general-purpose datasets (e.g., FineWeb-Edu, SlimPajama) at bespoke ratios without pre-baked replay constraints.
• Cybersecurity Text Evolution & Corpus Linguistics: Track the emergence of domain-specific terminology, public security disclosures, and changing community practices across more than ten years of web discourse.

Languages

The dataset consists primarily of English text filtered from the FineWeb Common Crawl archive.

---

Dataset Structure

Data Instances

The dataset is partitioned into 40 compressed JSON Lines files (00000.jsonl.gz to 00039.jsonl.gz). Each JSON record represents a unique deduplicated document identified as cybersecurity-relevant.

Data Fields

The dataset contains the following 12 fields per record:

• probability (float): Confidence score generated by the ModernBERT cybersecurity classifier.
• relevant (bool): Binary classification flag (true/false) confirming domain relevance.
• text (string): Full extracted text of the document.
• id (string): Unique Common Crawl URN document identifier (e.g., <urn:uuid:...>).
• dump (string): Source Common Crawl snapshot identifier (e.g., CC-MAIN-2024-51).
• url (string): Source web URL.
• date (string): Crawl / publication timestamp string.
• file_path (string): Source path within the FineWeb archive.
• language (string): Detected language tag (e.g., en).
• language_score (float): Confidence score from the language identifier.
• token_count (int): Document token count.
• idx (int): Record index identifier.

Files & Storage Summary

The corpus is packaged into 40 gzipped JSONL files totaling approximately 88.8 GB compressed:

File Range	Shard Count	Total Compressed Size	Format
00000.jsonl.gz – 00039.jsonl.gz	40 shards	~88.8 GB	Gzip-compressed JSONL (.jsonl.gz)

---

Usage

Loading via datasets

from datasets import load_dataset

# Load a specific shard
shard_0 = load_dataset(
    "RISys-Lab/RedSage-CFW-Full-Dedup",
    data_files="00000.jsonl.gz",
    split="train"
)
print(shard_0[0])

# Stream across the entire deduplicated dataset
full_stream = load_dataset(
    "RISys-Lab/RedSage-CFW-Full-Dedup",
    split="train",
    streaming=True
)

for sample in full_stream:
    print(sample["id"], sample["token_count"])
    break

Processing Directly in Python

import gzip
import json

with gzip.open("00000.jsonl.gz", "rt", encoding="utf-8") as f:
    for line in f:
        doc = json.loads(line)
        print(doc["id"], doc["probability"])
        break

---

Dataset Creation

Curation Rationale

Continual pretraining on redundant web documents can degrade training efficiency and trigger verbatim memorization. While RedSage-CFW Full (Non-Dedup) provides the complete raw classifier outputs for cross-crawl auditing, this deduplicated release provides a cleaner, computation-ready base across the entire 2013–2024 time span.

Source Data

• Base Corpus: HuggingFaceFW/fineweb, aggregated across 103 Common Crawl snapshots between 2013 and 2024 (~17.2T tokens).

Data Processing Pipeline

1. Classifier Filtering: A binary classifier based on ModernBERT-base (achieving 97.3% validation accuracy) was run across 103 FineWeb dumps, retrieving ~122.15M cybersecurity-relevant documents (~89.8B tokens).
2. Global Deduplication: MinHash-LSH was executed across the corpus using DataTrove, identifying and removing duplicate and near-duplicate text. This reduced total token volume by ~47.9%.
3. Partitioning: The deduplicated records were serialized into 40 gzip-compressed JSON Lines shards (.jsonl.gz).

---

Considerations for Using the Data

Social Impact

This dataset provides an accessible, non-redundant training resource for developing open-source cybersecurity foundational models, transparent assistants, and automated defensive tooling.

Discussion of Biases and Limitations

• Source Biases: As with all web-scale scrapes, the dataset contains informal syntax, forum noise, and potential biases from public discussions.
• Dual-Use Considerations: Contains authentic technical discussions, exploit mechanics, and vulnerability disclosures. Safe deployment requires post-training alignment and guardrails.

---

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}
}

Acknowledgements

Global deduplication was conducted using the open-source datatrove library.