This is a subset of rawrr_v2 that has chosen responses that are not chat completions as in continuing the prompt but actual responses. It might be a good idea to repeat training on this part of the dataset after previous DPO + SFT training to bring up uncensored-ness to the top, but without making the model dry as it happens with using unalignment/toxic-dpo-0.1 - I plant to test exactly that, hence the repo.
