metadata

license: mit
task_categories:
  - question-answering
language:
  - en
size_categories:
  - 10K<n<100K
pretty_name: ARD

AI Alignment Research Dataset

This dataset is based on alignment-research-dataset.

For more information about the dataset, have a look at the paper or LessWrong post.

It is currently maintained and kept up-to-date by volunteers at StampyAI / AI Safety Info.

Sources

The important thing here is that not all of the dataset entries contain all the same keys.

They all have the keys: id, source, title, text, and url

Other keys are available depending on the source document.

source: indicates the data sources:

agentmodels
aiimpacts.org
aipulse.org
aisafety.camp
arbital
arxiv_papers
audio_transcripts
carado.moe
cold.takes
deepmind.blog
distill
eaforum
gdocs
gdrive_ebooks
generative.ink
gwern_blog
intelligence.org
jsteinhardt
lesswrong
markdown.ebooks
nonarxiv_papers
qualiacomputing.com
reports
stampy
vkrakovna
waitbutwhy
yudkowsky.net

alignment_text: This is label specific to the arXiv papers. We added papers to the dataset using Allen AI's SPECTER model and included all the papers that got a confidence score of over 75%. However, since we could not verify with certainty that those papers where about alignment, we've decided to create the alignment_text key with the value "pos" when we manually labeled it as an alignment text and "unlabeled" when we have not labeled it yet. Additionally, we've only included the text for the "pos" entries, not the "unlabeled" entries.

Usage

Execute the following code to download and parse the files:

from datasets import load_dataset
data = load_dataset('StampyAI/alignment-research-dataset')

To only get the data for a specific source, pass it in as the second argument, e.g.:

from datasets import load_dataset
data = load_dataset('StampyAI/alignment-research-dataset', 'lesswrong')

The various sources have different keys - the resulting data object will have all keys that make sense, with `None** as the value of keys that aren't in a given source. For example, assuming there are the following sources with the appropriate features:

source1

id
name
description
author

source2

id
name
url
text

Then the resulting data object with have 6 columns, i.e. id, name, description, author, url and text, where rows from source1 will have None in the url and text columns, and the source2 rows will have None in their description and author columns.

Limitations and bias

LessWrong posts have overweighted content on x-risk doom so beware of training or finetuning generative LLMs on the dataset.

Contributing

Join us at StampyAI.

Citing the Dataset

Please use the following citation when using our dataset:

Kirchner, J. H., Smith, L., Thibodeau, J., McDonnell, K., and Reynolds, L. "Understanding AI alignment research: A Systematic Analysis." arXiv preprint arXiv:2022.4338861 (2022).