Hamzonium/cringe-meter-v2

Cringe Meter v2

A fine-tuned Laya model that sorts a LinkedIn or X draft into one of 8 archetypes in a single pass: genuine, humblebrag, fake parable, engagement bait, hustle guru, buzzword salad, shameless plug and AI thread bro. The Cringe Meter code shows 1 minus P(genuine) as a thermometer.

Code, extension and setup: https://github.com/Hamza29199/cringe-meter

Use

git clone https://github.com/Hamza29199/cringe-meter && cd cringe-meter
pip install laya torch huggingface_hub
hf download Hamzonium/cringe-meter-v2 --local-dir out/cringe_v2
python server.py

How it was trained

Only Laya's 2-layer decision head was trained, on a frozen mmBERT-base encoder, using about 6,200 synthetic posts (templates written with Claude). It runs in about 250 ms per post on a laptop CPU.

Limits

This is a prototype. The training data is synthetic and the test posts have a single author, so treat the numbers as a rough guide:

Fresh test set (64 posts)	Result
Accuracy over 8 archetypes	81% (chance 12.5%)
Cringe-score AUC	0.96
Genuine posts rated above 50% cringe	25%

Expect false alarms on real posts, for example sincere hiring posts. It works in English only.

Credits

Built on Laya by Convai Innovations (Apache-2.0) and the mmBERT-base encoder from JHU CLSP; check the encoder's own licence before reuse.