license: mit
dataset_info:
features:
- name: prompt
dtype: string - name: chosen
dtype: string - name: rejected
dtype: string - name: rationale
dtype: string
splits: - name: train
num_bytes: 35793725
num_examples: 12000 - name: test
num_bytes: 2545516
num_examples: 854
download_size: 21399617
dataset_size: 38339241
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: test
path: data/test-*
tags:
- split: train
- preference learning
- dpo
- rdpo
- rationale
An rationale-enhanced version of the paired preference learning dataset Intel-ORCA-DPO.
These rationales are general, high-level explanation of why the chosen response is preferred over the rejected response.
The dataset was generated according to this paper: Data-Centric Human Preference Optimization with
Rationales.