nova-8b-cybersec

Fine-tuned Dolphin3.0-Llama3.1-8B for cybersecurity tasks.

Model Details

• Base Model: cognitivecomputations/Dolphin3.0-Llama3.1-8B
• Fine-tuning: QLoRA (rank 64, alpha 128)
• Training Examples: 40,075
• Context Length: 8192 tokens
• Format: ChatML

Training Data

Dataset	Examples
SecurityGPT	16,000
PKI Context QA	16,278
Document Summaries	2,720
Elbranschen Threats	3,386
ISO 27001 Controls	1,116
ISO 27005 Threats	576

Usage

Ollama

ollama run pki/nova-8b-cybersec

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("pki/nova-8b-cybersec")
tokenizer = AutoTokenizer.from_pretrained("pki/nova-8b-cybersec")

Files

• model-*.safetensors - Model weights (4 shards)
• dolphin3-8b-nova.gguf - GGUF format for Ollama/llama.cpp
• tokenizer.json - Tokenizer

Training Config

• Epochs: 5
• Batch size: 2 (effective 40 with gradient accumulation)
• Learning rate: 5e-5
• LoRA rank: 64, alpha: 128
• Hardware: RTX 3090 24GB

License

Apache 2.0