nesilabs/tosilos-24b
nesilabs/tosilos-24b is a 24 billion parameter language model based on the Mistral Devstral 24B architecture, fine-tuned for European cybersecurity tasks. It is optimized for local deployment on a single 32GB GPU and demonstrates improved performance on cybersecurity metrics compared to its base model. This model is specifically designed for authorized security testing, research, and education.
Loading preview...
Overview
nesilabs/tosilos-24b is a 24 billion parameter language model, fine-tuned from the Mistral Devstral 24B base model. It is specifically designed for European cybersecurity applications and is optimized for local execution on a single 32GB GPU. The model was trained using QLoRA (r=64, alpha=128) on a curated Tosilos corpus of 17,000 domain-specific examples, alongside 25% general replay data.
Key Capabilities
- Enhanced Cybersecurity Performance: Achieves 92.8% on the CyberMetric benchmark, a 1.2 percentage point improvement over the Devstral 24B base model.
- General Knowledge Retention: Maintains general knowledge performance, scoring 77.4% on MMLU, indicating no significant forgetting during specialized fine-tuning.
- Domain-Specific Reasoning: Shows a modest but consistent improvement in domain-specific judgment tasks, as evaluated by Opus 5.
- Resource Efficient: Designed to run efficiently in 4-bit quantization on a single 32GB-class GPU, making it suitable for home server deployment.
Intended Use Cases
- Authorized Security Testing: Ideal for penetration testing and vulnerability assessment against systems where explicit, written authorization is granted.
- Cybersecurity Research: Supports academic and industry research into advanced security techniques and model behavior in cybersecurity contexts.
- Educational Purposes: Useful for training and educational programs focused on cybersecurity, providing a specialized tool for learning and experimentation.
Important Considerations
This model is provided "as is" and is intended strictly for authorized security work. Users are solely responsible for its deployment and must comply with all applicable laws and regulations.