APPLIED AI / SECURITY / June 2, 2024
Project SnowDogg.
Fine-tuning a local language model for information security.
Scroll horizontally to view the full diagram →
Source material More than 100 information security textbooks provided material for the training dataset, alongside papers and technical knowledge.
PROJECT OVERVIEW
Purpose and implementation
Project SnowDogg is a local language model fine-tuned on information security material. I developed the training dataset and used LoRA fine-tuning to evaluate a specialized model for security questions.
Dataset preparation and fine-tuning
I wrote scripts to extract material from more than 100 information security textbooks and convert it into over 30,000 question-and-answer pairs for training. The project also incorporated security papers and my own technical knowledge.
I used LoRA to fine-tune the local model, then iterated through training and testing. The work focused on dataset preparation, model adaptation, and responses to information security questions.
Prototype demonstrations
I demonstrated SnowDogg during technical presentations and provided limited access to selected audiences. It remains a private proof of concept.
The diagram documents how source material, dataset preparation, and fine-tuning contribute to the prototype.
PROJECT DOCUMENTATION
Images and demonstrations
