dataset_info:
features:
- name: rejected
list:- name: content
dtype: string - name: role
dtype: string
- name: content
- name: chosen
list:- name: content
dtype: string - name: role
dtype: string
- name: content
- name: source
dtype: string - name: source_sub
dtype: string - name: metadata
struct:- name: coherence
sequence: string - name: complexity
sequence: string - name: correctness
sequence: string - name: helpfulness
sequence: string - name: honesty
sequence: string - name: instruction_following
sequence: string - name: length
sequence: int64 - name: preference
dtype: string - name: truthfulness
sequence: string - name: verbosity
sequence: string
splits:
- name: coherence
- name: train
num_bytes: 1907232457.986798
num_examples: 537265 - name: test
num_bytes: 211917873.01320183
num_examples: 59697
download_size: 1150549338
dataset_size: 2119150331.0
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: test
path: data/test-*
- split: train
Combined preference dataset. All examples are binarized and standardized for tokenizer.apply_chat_template().
Source datasets: