Dataset Card for UltraFeedback Cleaned
Dataset Description
This is a cleaned version of the HuggingFaceH4/ultrafeedback_binarized
and was turned into jsonl format for DPO or PPO training.
I did the following clean steps:
Remove all lines with 'translation' or 'translate'. I believe few translation tasks are not good for fine-tuning.
Remove all answers starts with 'User: As an AI assistan'. It's a mistake that assistant answers have prompt.
Remove all lines with 'As an… See the full description on the dataset page: https://huggingface.co/datasets/flyingfishinwater/ultrafeedback_clean.
