alignment-research-dataset.py

import json
from pathlib import Path

import datasets
from datasets import Value, Sequence, Features


_CITATION = '''
@article{kirchner2022understanding,
  title={Understanding AI Alignment Research: A Systematic Analysis},
  author={Kirchner, Jan H and Smith, Logan and Thibodeau, Jacques and McDonnell, Kyle and Reynolds, Laria},
  journal={arXiv preprint arXiv:2022.4338861},
  year={2022}
}
'''

_DESCRIPTION = """The AI Alignment Research Dataset is a collection of documents related to AI Alignment and Safety from various books, research papers, and alignment related blog posts."""

_HOMEPAGE = "https://github.com/StampyAI/alignment-research-dataset"

_LICENSE = "MIT license"

_VERSION_ = '0.0.0'


def iterate_file(filename):
    print(filename)
    with open(filename) as f:
        for l in f:
            try:
                yield json.loads(l)
            except Exception as e:
                print(f'Could not parse: {l}')


## Feature extractor helpers
def get_type(value):
    """Recursively get the huggingface type for the provided value."""
    if value is None:
        return None
    if value and isinstance(value, (tuple, list)):
        return features.Sequence(
            get_type(value[0])
        )
    if value and isinstance(value, dict):
        return {k: get_type(v) for k, v in value.items()}
    if isinstance(value, str):
        return Value('string')
    if isinstance(value, int):
        return Value('int32')
    if isinstance(value, float):
        return Value('double')
    if isinstance(value, bool):
        return Value('bool')
    return None


def print_extra_features(files):
    """Go through all the provided files, and get the non default features for the given file.

    This can be done manually but would be a hassle.
    It's assumed that the files contain a json object on each line.
    """
    ignored_keys = [
        'comments',  # Comments are arbitrarily nested objects, which doesn't play nice with huggingface
    ]

    per_file = {}
    for filename in sorted(files):
        extra_types = {}
        for item in iterate_file(filename):
            for k, v in item.items():
                if (k not in extra_types or not extra_types[k]) and k not in ignored_keys and k not in DEFAULT_FEATURES:
                    extra_types[k] = get_type(v)
        per_file[filename] = extra_types

    print('DATASOURCES = {')
    for k, features in per_file.items():
        vals = ',\n'.join(f"        '{k}': {v}" for k, v in features.items())
        print(f"    '{k.stem}': #\n{vals}\n    $,".replace('#', '{').replace('$', '}'))
    print('}')


# These keys are present in all files
DEFAULT_FEATURES = {
    'id': Value('string'),
    'source': Value('string'),
    'title': Value('string'),
    'text': Value('large_string'),
    'url': Value('string'),
    'date_published': Value(dtype='string'),
    'authors': Sequence(feature=Value(dtype='string'), length=-1),
    'summary': Sequence(feature=Value(dtype='string'), length=-1),
    'source_type': Value(dtype='string'),
}


# Per datasource additional features
DATASOURCES = {
    'agentmodels': {
        'book_title': Value(dtype='string'),
    },
    'agisf': {},
    'aisafety.info': {},
    'alignmentforum': {
        'karma': Value(dtype='int32'),
        'votes': Value(dtype='int32'),
        'words': Value(dtype='int32'),
        'comment_count': Value(dtype='int32'),
        'tags': Sequence(feature=Value(dtype='string')),
        'modified_at': Value(dtype='string'),
    },
    'arbital': {
        'alias': Value(dtype='string'),
        'tags': Sequence(feature=Value(dtype='string')),
    },
    'arxiv': {
        'data_last_modified': Value(dtype='string'),
        'abstract': Value(dtype='string'),
        'author_comment': Value(dtype='string'),
        'journal_ref': Value(dtype='string'),
        'doi': Value(dtype='string'),
        'primary_category': Value(dtype='string'),
        'categories': Sequence(feature=Value(dtype='string'), length=-1),
    },
    'blogs': {
        'initial_source': Value(dtype='string'),
    },
    'distill': {
        'abstract': Value(dtype='string'),
        'journal_ref': Value(dtype='string'),
        'doi': Value(dtype='string'),
        'bibliography_bib': Sequence(feature={'title': Value(dtype='string')}, length=-1),
    },
    'eaforum': {
        'karma': Value(dtype='int32'),
        'votes': Value(dtype='int32'),
        'words': Value(dtype='int32'),
        'comment_count': Value(dtype='int32'),
        'tags': Sequence(feature=Value(dtype='string')),
        'modified_at': Value(dtype='string'),
    },
    'lesswrong': {
        'karma': Value(dtype='int32'),
        'votes': Value(dtype='int32'),
        'words': Value(dtype='int32'),
        'comment_count': Value(dtype='int32'),
        'tags': Sequence(feature=Value(dtype='string')),
        'modified_at': Value(dtype='string'),
    },
    'special_docs': {},
    'youtube': {},
}


def join_features(features, to_join):
    """Recursively join the provided dicts.

    `to_join` can either be a dict to be merged, or a list of dicts to merge.
    """
    if not to_join:
        return Features(features)
    if isinstance(to_join, dict):
        return Features(dict(features, **to_join))
    return join_features(dict(features, **to_join[0]), to_join[1:])


class AlignmentResearchDatasetConfig(datasets.BuilderConfig):
    """BuilderConfig for AlignmentResaerchDataset."""

    def __init__(self, sources, features, **kwargs):
        """BuilderConfig for AlignmentResaerchDataset.

        :param List[string] sources: the sources which will be used by this config
        """
        super().__init__(version=datasets.Version(_VERSION_), **kwargs)
        self.sources = sources
        self.features = join_features(DEFAULT_FEATURES, features)

    @property
    def files(self):
        return [f'{source}.jsonl' for source in self.sources]


class AlignmentResaerchDataset(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version(_VERSION_)

    BUILDER_CONFIGS = [
        AlignmentResearchDatasetConfig(
            name='all',
            description='All data files',
            sources=list(DATASOURCES.keys()),
            features=list(DATASOURCES.values())
        )
    ] + [
        AlignmentResearchDatasetConfig(name=source, sources=[source], features=features) for source, features in DATASOURCES.items()
    ]
    DEFAULT_CONFIG_NAME = 'all'

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=self.config.features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        downloaded_files = dl_manager.download_and_extract(self.config.files)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={'files': downloaded_files}
            )
        ]

    # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
    def _generate_examples(self, files):
        seen = set()

        def is_good(item):
            item_id = item and item.get('id')
            if not item_id or item_id in seen:
                return False
            seen.add(item_id)

            return item['text'] not in [None, '', 'n/a']

        def prepare_example(item):
            return item['id'], {k: item.get(k) for k in self.config.features}

        lines = (item for filename in files for item in iterate_file(filename))
        for item in map(prepare_example, filter(is_good, lines)):
            yield item