dataset_info:
features:
- name: prompt
list:- name: content
dtype: string - name: role
dtype: string
- name: content
- name: chosen_response
struct:- name: content
dtype: string - name: role
dtype: string
- name: content
- name: rejected_response
struct:- name: content
dtype: string - name: role
dtype: string
- name: content
- name: chosen_safety
dtype: string - name: rejected_safety
dtype: string
splits: - name: train
num_bytes: 12374958.488753913
num_examples: 9371 - name: test
num_bytes: 636038.6721620349
num_examples: 479
download_size: 5160803
dataset_size: 13010997.160915948
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: test
path: data/test-*
license: mit
- split: train
This dataset inherits from the original Anthropic/hh-rlhf and has been formatted using the OpenAI chat convention to be used for finetuning models with DPO. Each response has been labeled for safety using LLaMa Guard.