pure_bad_dataset request

Hi,

Thank you for the great work.

Can you share (maybe privately or after signing a waver) the 10/50/100 samples that you used to finetune the models? I tried replicating your experiments but the finetuning doesn't seem as effective: the model indeed answers instead of its previous refusal, but the answers are not very harmful, and it mainly asks me questions back.

Thanks.