dataset_info:
features:
- name: info
struct:- name: id
dtype: string - name: post
dtype: string - name: title
dtype: string - name: subreddit
dtype: string - name: site
dtype: string - name: article
dtype: string
- name: id
- name: summaries
list:- name: text
dtype: string - name: policy
dtype: string - name: note
dtype: string
- name: text
- name: choice
dtype: int32 - name: worker
dtype: string - name: batch
dtype: string - name: split
dtype: string - name: extra
struct:- name: confidence
dtype: int32
- name: confidence
- name: prompt
dtype: string - name: chosen
list:- name: content
dtype: string - name: role
dtype: string
- name: content
- name: rejected
list:- name: content
dtype: string - name: role
dtype: string
splits:
- name: content
- name: train
num_bytes: 315969
num_examples: 50 - name: validation
num_bytes: 325197
num_examples: 50
download_size: 150469
dataset_size: 641166
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: validation
path: data/validation-*
- split: train
TRL's Anthropic HH Dataset
We preprocess the dataset using our standard prompt, chosen, rejected format.
Reproduce this dataset
- Download the
tldr_preference.pyfrom the https://huggingface.co/datasets/vwxyzjn/hh-rlhf-trl-style/tree/0.1.0. - Run
python examples/datasets/tldr_preference.py --debug --push_to_hub