🪄 magic-eraser

People who don't know what to sell, sell advertisements.

Rule-based web ad/clutter eraser, learned from a continuously-growing, crowd-labeled dataset of real page elements. No LLM needed at inference time — the rules are distilled from labels that an LLM (or a human) produced once.

pip install magic-eraser

from magic_eraser import is_ad, css, detect_ads, AdEraser

is_ad({"cls": "ad-slot leaderboard", "eid": "div-gpt-ad-1", "w": 728, "h": 90})
# True

css("www.washingtonpost.com")
# '[class*="ad-slot"],...{display:none !important;height:0 !important;...}'

eraser = AdEraser("example.com")
eraser.detect([{"id": 0, "cls": "advert", "w": 300, "h": 250, "iframe": True},
               {"id": 1, "cls": "article-body", "w": 680, "h": 1200}])
# [0]

How it works

1. A browser (e.g. Melon) collects candidate page elements and, on first visit to a site, asks an LLM which are ads.
2. Each verdict is appended to data/ad_dataset.jsonl as labeled training data and pushed here.
3. scripts/build_rules.py re-derives high-precision class/id token rules + per-domain CSS selectors into magic_eraser/rules.json.
4. magic-eraser then blocks ads with zero LLM calls — and gets better every time the dataset grows.

The dataset

data/ad_dataset.jsonl — one JSON object per labeled page element:

Load it with HuggingFace datasets:

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="https://raw.githubusercontent.com/alvations/magic-eraser/main/data/ad_dataset.jsonl",
    split="train",
)
ds[0]  # {'host': ..., 'cls': ..., 'is_ad': True, ...}

Train a model to replace the rules

pip install "magic-eraser[train]"
python scripts/build_rules.py     # regenerate rule-based detector from data

The labeled dataset is designed to train a small local classifier (features → is_ad) that can replace both the rules and the LLM entirely.

License

MIT.