---
license: apache-2.0
language:
  - en
library_name: transformers
tags:
  - cybersecurity
  - soc
  - siem
  - mitre-attack
  - incident-response
  - threat-detection
  - security-operations
  - fine-tuned
  - qlora
  - unsloth
  - gguf
  - ollama
base_model: openai/gpt-oss-20b
model-index:
  - name: rhythmai-cybersec-20b
    results:
      - task:
          type: text-generation
          name: Cybersecurity Q&A
        metrics:
          - type: eval_loss
            value: 0.5773
            name: Validation Loss
          - type: train_loss
            value: 0.4873
            name: Training Loss
datasets:
  - AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.0
  - Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset
pipeline_tag: text-generation
---

# RhythmAI Cybersec 20B

A cybersecurity-specialized language model fine-tuned from [OpenAI GPT-OSS-20B](https://huggingface.co/openai/gpt-oss-20b) for Security Operations Center (SOC) tasks including alarm investigation, threat analysis, MITRE ATT&CK mapping, incident response, and log analysis.

Built for [RhythmAI](https://github.com/SyedCode01) -- an AI-powered SOC platform that integrates with LogRhythm SIEM.

## Model Details

| Property | Value |
|----------|-------|
| **Base Model** | [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) (MoE, 21B total / 3.6B active params) |
| **Architecture** | Mixture of Experts (MoE) with MXFP4 native quantization |
| **Fine-tuning Method** | QLoRA (4-bit) via [Unsloth](https://github.com/unslothai/unsloth) |
| **LoRA Rank** | 32 |
| **LoRA Alpha** | 64 |
| **LoRA Dropout** | 0.05 |
| **Target Modules** | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| **Training Precision** | 4-bit QLoRA with BF16 compute |
| **Context Length** | 65,536 tokens (64K) |
| **GGUF Format** | MXFP4 (13 GB) |
| **License** | Apache 2.0 (inherited from GPT-OSS) |

## Training Data

Fine-tuned on **9,702 curated cybersecurity examples** sourced from **137,122 raw examples** across 4 public datasets, aggressively filtered for SOC/SIEM relevance (7.1% acceptance rate):

| Source | Raw Size | After Filtering | Description |
|--------|----------|-----------------|-------------|
| [Fenrir v2.0](https://huggingface.co/datasets/AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.0) | 83,920 | ~5,000 | General cybersecurity Q&A |
| [Trendyol Cybersecurity](https://huggingface.co/datasets/Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset) | 53,202 | ~5,000 | Instruction-tuned cybersecurity |

**Filtering pipeline**: Keyword relevance scoring (minimum 2 matches from 60+ SOC-relevant terms), response length between 50-15,000 characters, MD5-based deduplication. Average response length: **2,627 characters (~656 tokens)**.

**Split**: 9,217 train (95%) / 485 validation (5%)

**Format**: OpenAI-compatible chat format:
```json
{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
```

### Cybersecurity Content Breakdown

#### MITRE ATT&CK Coverage

The training data references **424 unique MITRE ATT&CK technique IDs** across all 14 tactics:

| Tactic | Examples | Coverage |
|--------|----------|----------|
| Execution | 3,004 | 31.0% |
| Lateral Movement | 2,427 | 25.0% |
| Impact | 1,949 | 20.1% |
| Privilege Escalation | 1,637 | 16.9% |
| Persistence | 1,568 | 16.2% |
| Exfiltration | 1,425 | 14.7% |
| Defense Evasion | 1,277 | 13.2% |
| Collection | 1,080 | 11.1% |
| Reconnaissance | 900 | 9.3% |
| Discovery | 889 | 9.2% |
| Initial Access | 807 | 8.3% |
| Command and Control | 208 | 2.1% |
| Credential Access | 169 | 1.7% |
| Resource Development | 12 | 0.1% |

**Most referenced techniques**: T1078 (Valid Accounts, 1,451 examples), T1055 (Process Injection, 1,120), T1021 (Remote Services, 582), T1071 (Application Layer Protocol, 541), T1027 (Obfuscated Files, 378), T1566 (Phishing, 378), T1059 (Command and Scripting Interpreter, 376), T1562 (Impair Defenses, 339), T1203 (Exploitation for Client Execution, 323), T1041 (Exfiltration Over C2, 322).

#### Attack Types & Threat Categories

| Attack Type | Examples | Coverage |
|-------------|----------|----------|
| Phishing & Social Engineering | 9,546 | 98.4% |
| Remote Code Execution | 5,620 | 57.9% |
| Lateral Movement | 2,427 | 25.0% |
| Privilege Escalation | 1,637 | 16.9% |
| PowerShell-based Attacks | 731 | 7.5% |
| Supply Chain Attacks | 653 | 6.7% |
| Credential Dumping (Mimikatz/LSASS) | 393 | 4.1% |
| Insider Threats | 376 | 3.9% |
| Zero-Day Exploits | 375 | 3.9% |
| Man-in-the-Middle | 294 | 3.0% |
| Brute Force / Credential Stuffing | 264 | 2.7% |
| C2 Communication | 228 | 2.4% |
| DDoS / Denial of Service | 217 | 2.2% |
| Backdoors | 203 | 2.1% |
| Rootkits | 180 | 1.9% |
| SQL Injection | 177 | 1.8% |
| Buffer Overflow | 144 | 1.5% |
| Cross-Site Scripting (XSS) | 127 | 1.3% |
| Fileless Malware | 116 | 1.2% |
| Living Off The Land (LOLBins) | 80 | 0.8% |
| DNS Tunneling | 57 | 0.6% |

#### Log Source & SIEM Knowledge

| Log Type | Examples | Coverage |
|----------|----------|----------|
| Windows Event Logs (Event IDs) | 977 | 10.1% |
| Network Flow (NetFlow/PCAP) | 410 | 4.2% |
| IDS/IPS Alerts | 364 | 3.8% |
| Authentication Logs | 289 | 3.0% |
| Firewall Logs | 150 | 1.5% |
| DNS Logs | 123 | 1.3% |
| Syslog | 112 | 1.2% |

**Security platforms referenced**: Nmap (214), YARA rules (158), Microsoft Sentinel (120), Elastic/ELK (107), Wireshark (104), Splunk (70), Metasploit (65), Sigma rules (50), Snort/Suricata (45).

#### Compliance & Regulatory Frameworks

| Framework | Examples | Coverage |
|-----------|----------|----------|
| NIST (CSF/SP 800-series) | 9,620 | 99.2% |
| GDPR | 411 | 4.2% |
| HIPAA | 310 | 3.2% |
| OWASP | 304 | 3.1% |
| PCI-DSS | 152 | 1.6% |
| CIS Controls | 66 | 0.7% |
| ISO 27001 | 57 | 0.6% |
| SOC 2 | 35 | 0.4% |

## Training Details

| Parameter | Value |
|-----------|-------|
| **GPU** | NVIDIA RTX PRO 6000 Blackwell (96 GB VRAM) |
| **Framework** | Unsloth 2026.3.3 + Transformers 5.2.0 |
| **Epochs** | 3 |
| **Effective Batch Size** | 8 (2 per device x 4 gradient accumulation) |
| **Learning Rate** | 2e-4 (cosine schedule, 5% warmup) |
| **Optimizer** | AdamW 8-bit |
| **Weight Decay** | 0.01 |
| **Max Sequence Length** | 4,096 (training) / 65,536 (inference) |
| **Packing** | Enabled (short examples packed together) |
| **Gradient Checkpointing** | Unsloth optimized (30% VRAM savings) |
| **Total Steps** | 3,459 |
| **Training Time** | ~12.5 hours |
| **Trainable Parameters** | 67M / 21B (0.32%) |

## Training Metrics

| Metric | Value |
|--------|-------|
| **Final Training Loss** | 0.4873 |
| **Final Validation Loss** | 0.5774 |
| **Best Validation Loss** | 0.5773 (step 3,000) |
| **Initial Validation Loss** | 0.7866 (step 100) |

The model shows consistent improvement across training with no signs of overfitting (validation loss closely tracks training loss).

## Capabilities

This model is specialized for:

- **Alarm Investigation**: Analyzing security alarms from SIEM platforms with contextual threat assessment
- **MITRE ATT&CK Mapping**: Identifying tactics, techniques, and procedures (TTPs) from security events
- **Incident Response**: Generating structured incident response playbooks and triage recommendations
- **Threat Analysis**: Assessing threat severity, identifying indicators of compromise (IOCs)
- **Log Analysis**: Interpreting Windows Event Logs, firewall logs, IDS/IPS alerts, and authentication logs
- **Detection Engineering**: Suggesting detection rules and correlation logic
- **Compliance Guidance**: NIST, PCI-DSS, HIPAA, GDPR security control recommendations

## Usage

### With Ollama (Recommended)

```bash
# Create the model from GGUF
ollama create rhythmai-cybersec-20b -f Modelfile

# Run interactively
ollama run rhythmai-cybersec-20b "Analyze this security event: Multiple failed RDP login attempts from IP 203.0.113.45 targeting the domain controller, followed by a successful login and immediate PowerShell execution."

# Use via API
curl http://localhost:11434/api/chat -d '{
  "model": "rhythmai-cybersec-20b",
  "messages": [
    {"role": "system", "content": "You are a senior SOC analyst. Analyze security events and provide actionable recommendations."},
    {"role": "user", "content": "What MITRE ATT&CK techniques are associated with credential dumping?"}
  ]
}'
```

### With Transformers + PEFT (Adapter)

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", device_map="auto")
model = PeftModel.from_pretrained(base_model, "SyedCode01/rhythmai-cybersec-20b", subfolder="adapter")
tokenizer = AutoTokenizer.from_pretrained("SyedCode01/rhythmai-cybersec-20b", subfolder="adapter")

messages = [
    {"role": "system", "content": "You are a cybersecurity expert specializing in SOC operations."},
    {"role": "user", "content": "Explain lateral movement in the context of MITRE ATT&CK."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Example Output

**Prompt**: "What is lateral movement in cybersecurity?"

**Response**: Lateral movement in cybersecurity refers to the techniques attackers use to navigate through a network or system after gaining initial access. Once an attacker compromises a single device or account -- often with relatively low privileges -- they employ lateral movement to expand their reach, elevate their access levels, and ultimately achieve broader objectives such as data exfiltration or system control. This process involves moving from one compromised asset to another, leveraging existing network paths, user credentials, and administrative tools. Attackers might use legitimate remote management protocols (e.g., RDP, SSH), stolen passwords, or exploit vulnerabilities in software to traverse the environment. It's a critical phase in advanced persistent threat (APT) campaigns, as it enables attackers to remain stealthy over extended periods by using normal administrative functions and legitimate credentials.

## File Structure

```
rhythmai-cybersec-20b/
  adapter/                    # LoRA adapter weights (PEFT)
    adapter_config.json
    adapter_model.safetensors
    tokenizer.json
    tokenizer_config.json
  gguf-q4_k_m_gguf/          # GGUF for Ollama deployment
    gpt-oss-20b.MXFP4.gguf   # 13 GB quantized model
    Modelfile                 # Ollama model definition
  README.md                   # This model card
```

## Intended Use

This model is designed for cybersecurity professionals, SOC analysts, and security teams who need AI assistance with:
- Security alarm triage and investigation
- Threat intelligence analysis
- Incident response planning
- Security posture assessment
- MITRE ATT&CK framework mapping

## Limitations

- **Domain-specific**: Optimized for cybersecurity tasks; general knowledge may be less reliable than the base model
- **Not a replacement for human analysts**: Outputs should be validated by qualified security professionals
- **Training data bias**: Performance may vary for threats or attack patterns not well-represented in the training data
- **Context window**: Supports up to 65,536 tokens (64K); training used 4,096 max sequence length but the base model's full context capability is preserved
- **No real-time data**: The model does not have access to real-time threat intelligence feeds

## Citation

```bibtex
@misc{rhythmai-cybersec-20b,
  title={RhythmAI Cybersec 20B: A Fine-Tuned Cybersecurity Language Model},
  author={Syed Hasan Iqbal},
  year={2026},
  url={https://huggingface.co/SyedCode01/rhythmai-cybersec-20b},
  note={Fine-tuned from OpenAI GPT-OSS-20B for SOC operations}
}
```

## Acknowledgments

- [OpenAI](https://openai.com) for the GPT-OSS-20B base model (Apache 2.0)
- [Unsloth](https://github.com/unslothai/unsloth) for efficient QLoRA fine-tuning
- [AlicanKiraz0](https://huggingface.co/AlicanKiraz0) for the Fenrir v2.0 cybersecurity dataset
- [Trendyol](https://huggingface.co/Trendyol) for the cybersecurity instruction tuning dataset