RedSage: Cybersecurity LLM

• RedSage is an open-source, domain-specialized LLM designed for cybersecurity operations, combining domain-aware pretraining and agentic augmentation to enable expert-level interactions while preserving data privacy.
• It leverages a multi-stage data curation and augmentation process, using rigorous cybersecurity filtering, deduplication, and seed document strategies to build a high-quality, diverse training corpus.
• Evaluation shows RedSage outperforms baseline models by up to 5 points on cybersecurity and general benchmarks, ensuring strong performance in both specialized and general reasoning tasks.

RedSage is an open-source, domain-specialized LLM designed for cybersecurity operations and workflows. It addresses the challenge of supporting diverse, expert-level cybersecurity interactions without incurring the privacy and data protection risks of proprietary LLM APIs. RedSage utilizes a combination of domain-aware continual pretraining, massive agentic augmentation, and post-training alignment to achieve superior performance on closed-book, conversational, and general reasoning tasks. All models, benchmarks, and datasets are publicly released for on-premise deployment and reproducible research (Suryanto et al., 29 Jan 2026).

1. Domain Data Curation and Pretraining

RedSage leverages a multi-stage data curation approach, selectively filtering, deduplicating, and augmenting content from large-scale web corpora and authoritative resources.

CyberFineWeb Construction

• Corpus Source: The FineWeb corpus (Common Crawl 2013–2024; ∼17.2T tokens) serves as the base.
• Cybersecurity Filtering: A ModernBERT binary classifier, trained on 4.6M positive and negative samples, identifies cybersecurity-relevant content. This yields ∼125M candidate documents (89.8B tokens).
• Data Mixture Strategy: For continual pretraining, a 30% replay ratio of FineWeb-Edu (broad general knowledge, 1.3T tokens) is included in each chunk to mitigate catastrophic forgetting.
• Deduplication: Global MinHash-LSH (64-bit, 8 hashes, 14 buckets) removes 58% of documents and 48% of tokens, resulting in ∼52M documents (46.8B tokens).
• Final Selection: The last 5 chronological chunks are kept for cost control, yielding ∼13M documents (11.7B tokens).

RedSage-Seed Curation

• Authoritative Seed Documents: 28,637 Markdown files (0.15B tokens) are manually curated and organized:
  • Knowledge–General: MITRE ATT&CK, CWEs, OWASP Top 10, Wikipedia, roadmap.sh
  • Knowledge–Frameworks: CAPEC, CWE, MITRE ATT&CK, OWASP
  • Skills–Offensive: CTF write-ups, HackTricks, Nullbyte, PayloadsAllTheThings
  • Tools–CLI: TLDR pages, Linux man pages
  • Tools–Kali: Official Kali Linux documentation

Data Mixing Statistics

Selection Stage	%Docs (orig. FineWeb)	%Tokens (orig. FineWeb)
Filtered cybersecurity	0.51%	0.52%
After deduplication + general mix	0.21%	0.27%
Final used (for pretraining)	0.053%	0.068%

This targeted curation yields a pretraining dataset with high domain relevance and broad coverage, balancing authoritative depth with general contextual knowledge.

2. Agentic Augmentation Pipeline

RedSage adopts a two-stage agentic pipeline—modeled after AgentInstruct—for generating large-scale, realistic expert–assistant multi-turn dialogues across the cybersecurity domain.

Pipeline Components

1. Planner Agent:
   • Consumes Markdown input from RedSage-Seed (≤32K tokens per chunk).
   • Outputs structured JSON "plans," each covering skill sets (e.g., Network Discovery, Exploitation Workflows) with various “augmentation types” (e.g., ARP scan simulation, command-crafting).
   • Prompting enforces relevance, technical diversity, creativity, and conversational grounding.
2. Augmenter Agent:
   • Takes a single augmentation plan and original content.
   • Generates fixed-format, multi-turn (typically 8–12 turns) chat conversations. Each interaction features technical rigor, step-by-step reasoning, actionable best practices, and integrating references to standards (MITRE, NIST).

Augmentation Scale

Category	#Samples	Avg Tokens	Avg Turns
Knowledge–General	67,635	1,327	9.96
Knowledge–Frameworks	39,908	1,285	9.46
Skills–Offensive	38,870	1,345	10.10
Tools–CLI	109,261	1,331	9.73
Tools–Kali	10,506	1,356	9.26
Total	266,180	1,326	9.70

Agentic augmentation expands the supervised fine-tuning dataset by 9.2× in sample count and 2.3× in total tokens relative to seeds.

This methodology yields high-diversity, deeply contextualized training dialogues aligned with real-world expertise and operational workflows.

3. Model Architecture and Training Regimen

RedSage builds upon the Qwen3-8B-Base architecture, a Transformer-decoder LLM with 8B parameters. No architectural modifications are introduced.

Training Objectives and Procedure

• Pretraining Objective: Standard next-token prediction via cross-entropy:

$$\mathcal{L} = -\sum_{t=1}^T \sum_{v \in V} \mathbf{1}_{\{v = y_t\}} \log p_\theta(v \mid x_{<t})$$

Training Stages

1. Continued Pretraining (CPT)
   • Stage 1: CyberFineWeb (11.7B tokens); 1 epoch; batch size 1024 (on 32×A100); learning rate 2.5×10⁻⁶; AdamW optimizer; linear warmup (1K steps); ZeRO-3.
   • Stage 2: RedSage-Seed (0.15B tokens) plus RedSage-Dump (0.7B tokens); 1 epoch; same hyperparameters.
2. Supervised Fine-Tuning (SFT)
   • Data: RedSage-Conv (266K dialogues) + SmolTalk2 non-reasoning subset (~M samples).
   • Mixing Ratio: Implicitly ~15% cybersecurity-specific, 85% generic.
   • 2 epochs; batch size 1024; learning rate 2.5×10⁻⁵; cosine schedule; weight_decay=0.05; warmup=1%.
3. Direct Preference Optimization (DPO)
   • Data: Tulu 3 8B Preference Mixture (~250K preference pairs); α=0.2, β=1.0.

Total training compute: ~134 GPU-hours (4,096 GPU-h on 32×A100), including CPT, SFT, and DPO.

4. Evaluation Benchmarks and Results

RedSage's capabilities are assessed on both cybersecurity-specific and general LLM benchmarks, employing rigorous methodologies and head-to-head comparisons.

4.1 RedSage-Bench

A new evaluation suite, RedSage-Bench, comprises 30,000 multiple-choice questions and 240 open-ended QA items, equally distributed over Knowledge, Skills, and Tools. MCQs derive from Llama-3.3-70B Instructor prompts, validated via a two-stage LLM verifier. MCQs use normalized log-likelihood accuracy; open-ended items use LLM-as-Judge with a 0–10 quality rubric.

RedSage-Bench MCQ 0-shot Accuracy (8B Params)

Model	Macro	Gen	Frm	Off	CLI	Kali
Qwen3-8B-Base	84.24	83.08	81.94	88.23	85.08	78.86
RedSage-8B-CFW	84.86	83.62	83.30	88.81	85.30	79.32
RedSage-8B-Base	85.05	83.12	84.94	88.72	85.44	79.36
RedSage-8B-Ins (SFT)	85.73	84.20	84.98	89.06	86.80	80.30
RedSage-8B-DPO	84.83	82.48	83.80	88.54	86.30	79.30

SFT provides up to +1.49 points macro improvement over Qwen3-8B. On open-ended QA, DPO further boosts correctness (+7% absolute, 0.74→0.81) and quality score (+0.07, 6.43→6.50).

4.2 Standard Cybersecurity Benchmarks

Model	Mean	CTI-MCQ	RCM	CyMtc	MMLU-CSec	SecBench	SECURE-CWET
Qwen3-8B-Base	80.81	68.80	63.50	93.80	86.00	83.62	93.33
RedSage-8B-Base	84.56	71.04	78.40	92.60	87.00	81.76	93.22
RedSage-8B-DPO	81.10	70.84	70.60	90.00	79.00	80.06	91.35

Improvements of +3.75 (CPT) to +5.59 points (DPO) macro over the Qwen3-8B-Base.

4.3 General LLM Benchmarks

Accuracy on Open LLM Leaderboard (MMLU, ARC-C, GSM8K, HellaSwag, TruthfulQA, WinoGrande, IFEval):

Model	Mean	MMLU	ARC-C	GSM8K	HSwag	TQA	WinoG	IFEval
Qwen3-8B-Base	70.86	78.73	68.09	81.73	79.62	43.84	73.16	–
RedSage-8B-Ins (SFT)	73.34	77.38	69.62	86.05	79.00	47.75	73.64	79.97
RedSage-8B-DPO	74.33	77.07	71.76	82.71	79.87	52.47	73.01	83.44

RedSage variants outperform base models by 3.48 to 3.77 points mean accuracy.

Evaluation Methodology

RedSage's evaluation pipeline applies 0-shot or 5-shot prompting as standard for each benchmark. MCQ accuracy reflects direct answer prediction without few-shot context. Open-ended QAs leverage a rubric-based LLM judge for human-like assessment.

5. Implications and Discussion

RedSage empirically demonstrates that large-scale, domain-aware continual pretraining—spanning both broad filtered web data and manually curated seeds—confers synergistic gains: broad general coverage from web-scale corpora complements high-precision, framework-grounded expertise from selected documentation. Agentic augmentation on this seed corpus yields an order of magnitude more diverse, operationally grounded multi-turn chats, which improve both factual accuracy and conversational depth.

Supervised fine-tuning that integrates cybersecurity dialogues while maintaining a majority representation of generic instructions ensures RedSage sustains generalized reasoning and instruction-following capabilities, rather than catastrophic overspecialization. Direct Preference Optimization further improves helpfulness and free-form response quality.

A notable finding is that cybersecurity specialization does not degrade general language or reasoning capability; instead, it yields improvements of 3–5 points on standard LLM benchmarks such as MMLU, ARC-C, and GSM8K—a result suggesting that well-structured domain expertise and high-quality, agentic conversational data positively reinforce multi-step reasoning and answer clarity.

RedSage’s open-source release of all artifacts facilitates on-premise, privacy-preserving deployments suitable for security operations centers and research environments where confidentiality requirements preclude third-party cloud APIs. The reproducible data curation, augmentation, and training pipeline may inform future work on other specialized domains.