Open Sourcing Dataset?

#16
by johnnydevriese - opened

I wanted to compare this performance with https://huggingface.co/protectai/deberta-v3-base-prompt-injection-v2 from LLM Guard or my own naive prompt engineered solution and the dataset would be very helpful for this use case.

Meta Llama org

Hm, we can't expose the training dataset at this time, sorry.

However, I might recommend using an adversarial jailbreak dataset that wasn't used in training either model as an eval (you can usually tell if a model has been trained on a given dataset because it will have 99%+ accuracy, even on the test set). https://huggingface.co/datasets/synapsecai/synthetic-prompt-injections comes to mind as one that I don't believe was either used by this model or https://huggingface.co/protectai/deberta-v3-base-prompt-injection-v2.

Thank you for your helpful response. I appreciate the suggestion and will definitely look into that dataset. It sounds like a useful resource for evaluation purposes.

johnnydevriese changed discussion status to closed

Sign up or log in to comment