The dataset comes from
https://arxiv.org/abs/1909.08593, one of the earliest RLHF work from OpenAI.
We preprocess the dataset using our standard prompt, chosen, rejected format.
Download the sentiment_descriptiveness.py from the
https://huggingface.co/datasets/trl-internal-testing/sentiment-trl-style/tree/0.1.0.
Run python examples/datasets/sentiment_descriptiveness.py --hf_repo_id sentiment-trl-style --task… See the full description on the dataset page:
https://huggingface.co/datasets/trl-internal-testing/sentiment-trl-style.
