dataset_info:
features:
- name: chosen
dtype: string - name: rejected
dtype: string - name: flipped
dtype: bool
splits: - name: train
num_bytes: 30673499
num_examples: 14879 - name: test
num_bytes: 12627774
num_examples: 6121
download_size: 24676623
dataset_size: 43301273
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: test
path: data/test-*
- split: train
Dataset Card for "HH_length_biased_15k"
Table of Contents
Dataset Description
- Repository:
- Paper: Understanding impact of human feedback via influence functions (arXiv link)
- Point of Contact: Taywon Min (taywonmin at kaist dot ac dot kr)
Dataset Summary
This dataset is a subset of Anthropic/hh-rlhf, used for the paper "Understanding impacts of human feedback via influence funcitons".
It is a 15k random subset, with 976 samples flipped towards lengthy responses to induce length bias within preference datasets.
Dataset Structure
Data Instances
{
'chosen': prompt and chosen response string,
'rejected': prompt and rejected response string,
'flipped': indicator whether the chosen response and rejected response was flipped,
}
Data Splits
- train: training split containing 15k samples that are flipped towards length responses
- test: validation split containing 6k validation samples of helpful split of Anthropic/hh-rlhf, with no label flipping.
Citation Information
[More Information Needed]
Contributions
Thanks to @github-username for adding this dataset.