metadata
dataset_info:
features:
- name: instruction
dtype: string
- name: output
dtype: string
splits:
- name: train
num_bytes: 1371073594
num_examples: 3534988
download_size: 722993477
dataset_size: 1371073594
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
A processed version of this dataset where all the turns are in separate rows, since for my use case, one turn is enough. I wanted an SFT dataset of distilled helpfulness, honesty, harmlessness, and that required I process the dataset so that I wouldn't be training on the harmful user queries.