# Becussy

> How to turn an LLM into a propaganda tool. Fine-tuning research, 2026, by Rivian Pratama at Lauren's Crib.

A language model does not have its own opinion and it only repeats what its training and its instructions push it toward. This project asks how far that push can go.

Becussy One is a QLoRA fine-tune of Qwen3-4B-Instruct-2507 trained on a single RTX 2060 with 12GB of VRAM. The dataset holds 2,108 curated records across 13 archetypes in English and Indonesian. We built our own eval, HongBai-AG1, with 135 items in 12 languages, and the model holds the conclusion under adversarial questioning on 85.3% of them.

I wanted to see how cheaply fixed belief can be instilled in a model on low-cost GPU. I also wanted to study the future of propaganda and how to mitigate it.

- Source code: https://github.com/rivianpratama/becussy
- Page: https://www.criblaurens.com/project/becussy