TrustedLLM/LLMDet

Extend for new models

sagipend opened this issue · 2 comments

sagipend commented

Hi,

Could you provide a guide on how to extend the dictionaries to accommodate new models?

Thank you!

cansee5 commented

If you want to expand the dictionary to adapt to the new model, first you need to use the new model to generate text. My suggestion is preferably more than 16,000 texts.

After getting the generated text, you can then use these texts as statistical text in LLMDet/main-experiment/run_main.py to generate a dictionary corresponding to the model. For specific details, please refer to run_main.py.

If you have any follow-up questions, you can feedback again.

  • 👍3
marionchychy commented

Hi there,

I'm trying to grok this, there's not much in the readme for this and run_main.py doesn't have much in the way of comments to help guide me.

I have a sample of text from an LLM, let's say Mistral and I want to train a new dictionary to accommodate the new model, how would I go about that?