Asura

Inspiration

Maxime Labonne's article on uncensoring llms https://huggingface.co/blog/mlabonne/abliteration

What it does

This framerwork "adds" to refusal paths of existing LLMs, making them believe in a new ground truth. Hence making them immune to prompt injection attacks

How we built it

The concept of abliteration is pre-existing. This demo uses an abliterated gemma model which is run as a llama-index agent with brightdata mcp tool.

Challenges we ran into

Spent a lot of time in using decompute apis only to find out they didn't work.

Accomplishments that we're proud of

refusal script works great on existing models!

What we learned

need better time management

What's next for Asura

publish and help community learn!

Built With

• brightdata
• gemma-2
• llamaindex
• python