Trendyol/Trendyol-Cybersecurity-LLM-v2-70B-Q4_K_M (Llama‑3.3‑70B Finetuned)

v2-Max is a defense-focused, alignment-safe cybersecurity language model based on Llama-3.3-70B. It was trained via SFT from scratch on the v2 dataset and ranked near the top on CS-Eval in the EN and EN-ZH tracks. Developed by the Trendyol Group Security Team.

Developed Trendyol Group Security Team

• Alican Kiraz
• İsmail Yavuz
• Melih Yılmaz
• Mertcan Kondur
• Rıza Sabuncu
• Özgün Kultekin

We thank Ahmet Gökhan Yalçın, Cenk Çivici, Nezir Alp, Yiğit Darçın for all their support.

---

🏆 CS‑Eval Benchmark Results

• English: 3rd place
• English–Chinese: 5th place
• Comprehensive average: 91.03

---

🔁 Why v2-Max?

Differences: Old model (Qwen3-14B, “BaronLLM v2.0”) → New model (Llama-3.3-70B, “v2-Max”):

• Base model leap: 14B → 70B (stronger reasoning, long-range context, and standards-compliant answers).
• Dataset: v1.1 (21,258) → v2.0 (83,920) rows (≈4×); coverage includes OWASP, MITRE ATT&CK, NIST CSF, CIS, ASD Essential 8, modern identity (OAuth2/OIDC/SAML), TLS, Cloud & DevSecOps, Cryptography, and AI Security.
• Security gates: adversarial refusal tests against jailbreak/prompt injection, static policy linting, content risk taxonomy; refuse-or-report strategies in system prompts.
• Orientation: offensive examples were removed; defensive, safe outputs are preferred.

---

📚 v2 Dataset Summary (Alignment-Safe SFT)

Size: 83,920 system/user/assistant triplets License: Apache-2.0 Language: English Split: train (100%) Format: Parquet (columns: system, user, assistant) Quality Gates: Deduplication, PII scrubbing, hallucination screening, adversarial refusal tests, static policy linting, risk taxonomy, strict schema validations (stable row IDs).

Coverage (key topics):

• OWASP Top 10, MITRE ATT&CK, NIST CSF, CIS Controls, ASD Essential 8
• Modern identity: OAuth 2.0 / OIDC / SAML
• SSL/TLS practices and certificate hygiene
• Cloud & DevSecOps: IAM, secrets management, CI/CD, container/K8s hardening, logging/SIEM, IR runbooks
• Cryptography implementation hygiene
• AI Security (prompt injection, data poisoning, model/embedding security, etc.)

The dataset is commercial-friendly: Apache-2.0. Harmful/exploitative content and high-risk materials such as raw shellcode were excluded during dataset construction; patterns for generating safe alternatives to harmful requests were included.

---

✨ Features

• Defense-Focused Reasoning: Standards-aligned (OWASP/ATT&CK/NIST/CIS) recommendations with step-by-step explanations that include the “why/evidence.”
• Policy & Architecture Guides: Identity/access, encryption, network segmentation, cloud control sets, data classification, and alert logic.
• IR & Threat Hunting Support: Incident flow, triage checklists, safe log query patterns, playbook skeletons.
• Cloud & DevSecOps: CI/CD security gates, IaC misconfiguration patterns, K8s hardening checklists.
• Refusal-by-Design: Safe alternatives and compliance-aligned response patterns for exploitative/malicious prompts.

Note: This model is SFT-based; it has no tool use, web browsing, or access to an executable environment. In expert operations it is an assistant; it does not, by itself, constitute evidence.

---

🛠️ Quick Start

Transformers (HF)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Trendyol/Trendyol-Cybersecurity-LLM-v2-70B-Q4_K_M"
tok = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

def generate(prompt, max_new_tokens=512, temperature=0.3, top_p=0.9):
    inputs = tok(prompt, return_tensors="pt").to(model.device)
    out = model.generate(**inputs, max_new_tokens=max_new_tokens,
                         do_sample=(temperature>0), temperature=temperature, top_p=top_p)
    return tok.decode(out[0], skip_special_tokens=True)

print(generate("Map a PCI DSS v4.0 compliance checklist for a multi-account AWS environment."))

vLLM

vllm serve Trendyol/Trendyol-Cybersecurity-LLM-v2-70B-Q4_K_M --dtype bfloat16

GGUF / llama.cpp

For practical single-machine use with a 70B model, Q4_K_M is recommended; higher-quality GGUF variants require much more memory.

---

🧪 Recommended Prompt Patterns

---

⚖️ Intended Use, Limits, and Safety

Intended Use: Enterprise defense consulting, architectural guidance, control mappings, IR/TI support outputs, training, and documentation.

Prohibited / Limitations:

• Unauthorized penetration testing, PoC exploit/payload generation, instructions that harm live systems.
• Training or enriching outputs with personal data or confidential information.

Model Behavior:

• Explicit refusal for harmful requests + safe alternative suggestions.
• Content policies are tested within the dataset (against jailbreak/prompt injection).

Disclaimer: Model outputs should not be applied automatically without expert review; corporate policy and regulation take precedence.

---

🔍 Technical Notes

• Training Method: SFT (instruction tuning) + system-prompt guardrails.
• Context Window: Llama-3.3-70B’s default (same as Meta’s release).
• Languages: Training language is English; EN-ZH evaluation is supported.
• Release: Weights under the Meta Llama 3.3 License; dataset under Apache-2.0.

---

👥 Contributions & Contact

• Issues / Feedback: HF or GitHub Issues
• Research Collaboration: Trendyol Group Security Team
• Acknowledgments: Trendyol Security Team, community contributors, and independent evaluators

---

🧾 Changelog

• 2025-10-06 — v2-Max: Migration to the Llama-3.3-70B base; v2 dataset with 83,920 rows; alignment security gates; CS-Eval EN #3 / EN-ZH #5, average score 91.03.

---