Mini LLM Detector Benchmark — Starter

This is a tiny, runnable starter to build a 20–120 paragraph benchmark to test simple AI-text detectors.

What you’ll do

1. Fill data/data.csv with paragraphs (human + AI).
2. Install deps: pip install -r requirements.txt
3. Run: python src/eval.py
4. See: results/metrics.csv and results/summary.txt
5. Write your 1‑pager using report/one_pager_template.md

Data format (CSV)

Columns (keep these headers exactly): id,text,label,domain,paraphrase_level,language,source,notes

• label: human or ai
• domain: stem, policy, narrative (use these three to start)
• paraphrase_level: none, light, heavy
• language: en or es->en (if you machine-translate to Spanish and back)
• source: e.g., self, ku_website, wikipedia, chatgpt
• notes: short free text (e.g., prompt name, link slug)
• Wrap the text field in double quotes. If your text contains quotes, double them.

Quick test

A few sample rows are already in data/data.csv. Replace them with your own.