add loader script

README.md

@@ -17,7 +17,7 @@ It is currently maintained and kept up-to-date by volunteers at StampyAI / AI Sa
 
 ## Sources
 
-The important thing here is that not all of the dataset entries contain all the same keys. 
+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
 
@@ -55,6 +55,41 @@ Other keys are available depending on the source document.
 
 2. `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](https://coda.io/d/AI-Safety-Info_dfau7sl2hmG/Get-involved_susRF#_lufSr).
@@ -64,4 +99,3 @@ Join us at [StampyAI](https://coda.io/d/AI-Safety-Info_dfau7sl2hmG/Get-involved_
 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).
-

alignment-research-dataset.py

@@ -0,0 +1,315 @@
+import json
+from pathlib import Path
+
+import datasets
+from datasets import Value, Sequence
+
+
+_CITATION = '''
+@article{kirchner2022understanding,
+  title={Understanding AI Alignment Research: A Systematic Analysis},
+  author={Kirchner, Joshua H and Smith, Lauren and Thibodeau, Joseph and McDonnell, Kathleen and Reynolds, Lauren},
+  journal={arXiv preprint arXiv:2022.4338861},
+  year={2022}
+}
+'''
+
+_DESCRIPTION = """A dataset of AI alignment research, collected from various sources."""
+
+_HOMEPAGE = "https://github.com/StampyAI/alignment-research-dataset"
+
+_LICENSE = ""
+
+_VERSION_ = '0.0.0'
+
+
+def iterate_file(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'),
+}
+
+# Per datasource additional features
+DATASOURCES = {
+    'agentmodels': {
+        'source_filetype': Value(dtype='string', id=None),
+        'converted_with': Value(dtype='string', id=None),
+        'book_title': Value(dtype='string', id=None),
+        'authors': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None)
+    },
+    'aiimpacts.org': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+    'aipulse.org': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+    'aisafety.camp': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+    'alignment_newsletter': {
+        'converted_with': Value(dtype='string', id=None),
+        'source_type': Value(dtype='string', id=None),
+        'venue': Value(dtype='string', id=None),
+        'newsletter_category': Value(dtype='string', id=None),
+        'highlight': Value(dtype='int32', id=None),
+        'newsletter_number': Value(dtype='string', id=None),
+        'summarizer': Value(dtype='string', id=None),
+        'opinion': Value(dtype='string', id=None),
+        'prerequisites': Value(dtype='string', id=None),
+        'read_more': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None)
+    },
+    'arbital': {
+        'source_filetype': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None),
+        'alias': Value(dtype='string', id=None)
+    },
+    'arxiv_papers': {
+        'authors': Value(dtype='string', id=None),
+        'source_type': Value(dtype='string', id=None),
+        'converted_with': Value(dtype='string', id=None),
+        'data_last_modified': Value(dtype='string', id=None),
+        'abstract': Value(dtype='string', id=None),
+        'author_comment': Value(dtype='string', id=None),
+        'journal_ref': Value(dtype='string', id=None),
+        'doi': Value(dtype='string', id=None),
+        'primary_category': Value(dtype='string', id=None),
+        'categories': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None)
+    },
+    'audio_transcripts': {
+        'source_filetype': Value(dtype='string', id=None),
+        'converted_with': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None)
+    },
+    'carado.moe': {
+        'source_type': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None)
+    },
+    'cold.takes': {},
+    'deepmind.blog': {
+        'source_type': Value(dtype='string', id=None)
+    },
+    'distill': {
+        'source_type': Value(dtype='string', id=None),
+        'converted_with': Value(dtype='string', id=None),
+        'authors': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None),
+        'abstract': Value(dtype='string', id=None),
+        'journal_ref': Value(dtype='string', id=None),
+        'doi': Value(dtype='string', id=None),
+        'bibliography_bib': Sequence(feature={'title': Value(dtype='string', id=None)}, length=-1, id=None)
+    },
+    'eaforum': {
+        'authors': Value(dtype='string', id=None),
+        'score': Value(dtype='string', id=None),
+        'omega_karma': Value(dtype='string', id=None),
+        'votes': Value(dtype='string', id=None),
+        'tags': Value(dtype='string', id=None)
+    },
+    'gdocs': {
+        'source_filetype': Value(dtype='string', id=None),
+        'converted_with': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None),
+        'docx_name': Value(dtype='string', id=None)
+    },
+    'gdrive_ebooks': {
+        'source_filetype': Value(dtype='string', id=None),
+        'converted_with': Value(dtype='string', id=None),
+        'chapter_names': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None),
+        'file_name': Value(dtype='string', id=None)
+    },
+    'generative.ink': {},
+    'gwern_blog': {
+        'authors': Value(dtype='string', id=None)
+    },
+    'intelligence.org': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+    'jsteinhardt.wordpress.com': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+    'lesswrong': {
+        'authors': Value(dtype='string', id=None),
+        'score': Value(dtype='string', id=None),
+        'omega_karma': Value(dtype='string', id=None),
+        'votes': Value(dtype='string', id=None),
+        'tags': Value(dtype='string', id=None)
+    },
+    'markdown.ebooks': {
+        'source_type': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None),
+        'filename': Value(dtype='string', id=None)
+    },
+    'nonarxiv_papers': {
+        'source_filetype': Value(dtype='string', id=None),
+        'abstract': Value(dtype='string', id=None),
+        'authors': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None),
+        'filename': Value(dtype='string', id=None)
+    },
+    'qualiacomputing.com': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+    'reports': {
+        'source_filetype': Value(dtype='string', id=None),
+        'abstract': Value(dtype='string', id=None),
+        'authors': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None),
+        'filename': Value(dtype='string', id=None)
+    },
+    'stampy': {
+        'source_filetype': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None),
+        'question': Value(dtype='string', id=None),
+        'answer': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None),
+        'entry': Value(dtype='string', id=None)
+    },
+    'vkrakovna.wordpress.com': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+    'waitbutwhy': {
+        'source_type': Value(dtype='string', id=None),
+        'authors': Value(dtype='string', id=None)
+    },
+    'www.yudkowsky.net': {
+        'paged_url': Value(dtype='string', id=None)
+    },
+}
+
+
+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 datasets.Features(features)
+    if isinstance(to_join, dict):
+        return datasets.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):
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={'files': dl_manager.download(self.config.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))
+        return map(prepare_example, filter(is_good, lines))