metadata
dataset_info:
features:
- name: prompt
dtype: string
- name: chosen
list:
- name: content
dtype: string
- name: role
dtype: string
- name: rejected
list:
- name: content
dtype: string
- name: role
dtype: string
splits:
- name: train
num_bytes: 362466959
num_examples: 99717
- name: validation
num_bytes: 36867461
num_examples: 10555
- name: test
num_bytes: 38711045
num_examples: 10454
download_size: 109527109
dataset_size: 438045465
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
The binarized version of https://huggingface.co/datasets/stanfordnlp/SHP dataset ready to be used for DPO training The datset was binarized using the same procedure as in Human-Aware Loss Functions (HALOs)https://github.com/ContextualAI/HALOs/tree/main