This repo contains the general instruction-augmented corpora (containing 200M instruction-response pairs covering 40+ task categories) used in our paper Instruction Pre-Training: Language Models are Supervised Multitask Learners.
We explore supervised multitask pre-training by proposing Instruction Pre-Training, a framework that scalably augments massive raw corpora with instruction-response… See the full description on the dataset page:
https://huggingface.co/datasets/instruction-pretrain/general-instruction-augmented-corpora.
