Applied AI Build

Fine-tuning a compact model for phishing triage.

A fine-tuning workflow for adapting a compact language model to classify phishing-style messages and return analyst-ready summaries.

Use case Phishing review, analyst summaries, and quicker first-pass triage.

Workflow Dataset prep, LoRA fine-tuning, validation split, and evaluation.

Stack Python, Transformers, PEFT, prompt formatting, and metric review.

Output Training pipeline, sample dataset, evaluation script, and project notes.

Project view

I built a workflow for teaching a smaller AI model how to separate phishing-style messages from benign ones and return a short analyst-ready summary instead of a generic response.

Project access

Start with the project notes for the overview. Open the training script for the fine-tuning pipeline, the evaluation script for the review stage, and the sample dataset if you want to inspect the data format.

Project notes PDF Training script Evaluation script

Build Shape

How the fine-tuning workflow is structured.

01 Dataset prep

labelled examples and prompt structure

02 Base model

compact instruction model selected

03 LoRA fine-tuning

parameter-efficient adaptation

04 Evaluation

reviewing output quality and failure cases

Basic overview

This project fine-tunes a compact model on labelled phishing-style messages so it can separate suspicious content from benign content and return a short triage summary.

The aim was to build a full applied AI workflow rather than just use a model as-is: prepare the dataset, adapt the base model efficiently, evaluate the output, and produce something usable in an analyst-facing workflow.

Use case

The model is aimed at a practical workflow: separating phishing-like content from benign messages and returning a short summary that an analyst or reviewer can scan quickly.

What the workflow includes

Structured JSONL data with labels and target summaries.
Parameter-efficient fine-tuning with LoRA instead of full-model retraining.
Validation split and evaluation output to review performance and failure cases.
Training and evaluation scripts that can be extended to stronger datasets.

What this demonstrates

This project is here to show applied AI capability properly: model adaptation, data formatting, training workflow design, and evaluation rather than generic "AI interest" claims.