Nice and Great Work
Hey Sagar Verma, very wonderful and great work, I’m really impressed. Are you interested in collaborating with me? I also build AI models and was curious why you chose LLaMA specifically, since there are other models around 8B that can understand prompts better and give stronger explanations. Personally, I feel LLaMA lacks in understanding. But as of I see that the LLaMA have very less refusal rates and censorship than other models
Thanks @UJJAWAL-TYAGI for noticing the design choice. LLaMA v3 strikes a solid balance between reasoning and deployment efficiency. To uncensor an LLM, I use the abliteration method, leveraging activations from harmless vs. harmful prompts to find a refusal direction, then adjusting weights to stop refusals. This is often crucial for medical use cases. I’m also experimenting with another model with near-zero refusal rates for defense and cyber-security purpose, though it does raise ethical considerations.
Article: https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction
Also I’m always open to collaborating, feel free to ping me on LinkedIn or X anytime.