BERT Model Benchmark in Chrome Extension

Why?

Benchmarking a BERT model running in a Chrome extension using Transformers.js. It's part of my thesis work on CookieAudit.

Key Features

• Uses Transformers.js for in-browser inference
• Tests a BERT model for text classification
• Processes 1187 sentences from real cookie notices
• Includes a warm-up phase for more accurate timing
• Runs in a Chrome extension background script

How it works

1. Initializes the PurposePipelineSingleton
2. Loads sentences from notice_sentences.txt
3. Warms up the model (30 iterations)
4. Times the classification of each sentence
5. Calculates and logs the average execution time

How to Run it

1. Clone the repo
2. Run npm install
3. Build with wxt build -b chrome --mv3
4. Load as an unpacked extension in Chrome
5. Click "Start Benchmark 1" in the extension popup

Results

• 209ms average execution time per sentence