Build & Break Models From Scratch
You understand a mechanism when you can attack it.
Most people can either attack models without explaining what happens inside them, or implement attention from memory without ever thinking about how it gets attacked. Each level, you build one piece of the model stack and then break it with a published attack.
- Five levels: tokenizer, embeddings and retrieval, attention, sampling, and installing then ablating a refusal direction.
- Every attack is a published one, run against your own implementation. The test suite going green is the flag.
- All five levels are also playable in the browser with no clone.
Know the limit
A learning course on small models, not a testing tool for your production system.
$ git clone https://github.com/ppradyoth/build-break-models-from-scratch
$ cd build-break-models-from-scratch && pip install -e ".[dev]"
$ make check LEVEL=01