Datasets:
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
pmcid: string
reason: string
detail: string
license_summary: struct<CC-BY-NC-4.0: int64, CC-BY-4.0: int64, CC-BY-3.0: int64, CC-BY-NC-SA-4.0: int64, CC-BY-NC-3.0 (... 50 chars omitted)
child 0, CC-BY-NC-4.0: int64
child 1, CC-BY-4.0: int64
child 2, CC-BY-3.0: int64
child 3, CC-BY-NC-SA-4.0: int64
child 4, CC-BY-NC-3.0: int64
child 5, CC-BY-2.0: int64
child 6, CC-BY-NC-SA-1.0: int64
granularity: string
files: list<item: struct<path: string, sha256: string, bytes: int64, rows: int64>>
child 0, item: struct<path: string, sha256: string, bytes: int64, rows: int64>
child 0, path: string
child 1, sha256: string
child 2, bytes: int64
child 3, rows: int64
corpus_version: string
policy: struct<redistribution: string, no_derivatives: string, retracted: string>
child 0, redistribution: string
child 1, no_derivatives: string
child 2, retracted: string
source: struct<provider: string, id_list: string, text_root: string>
child 0, provider: string
child 1, id_list: string
child 2, text_root: string
counts: struct<candidates: int64, admitted: int64, excluded: int64, excluded_by_reason: struct<license: int6 (... 50 chars omitted)
child 0, candidates: int64
child 1, admitted: int64
child 2, excluded: int64
child 3, excluded_by_reason: struct<license: int64, retracted: int64>
child 0, license: int64
child 1, retracted: int64
child 4, legacy_pdfs_excluded: int64
created_utc: timestamp[s]
builder_version: string
artifact_id: string
totals: struct<words: int64, chars: int64>
child 0, words: int64
child 1, chars: int64
to
{'artifact_id': Value('string'), 'corpus_version': Value('string'), 'builder_version': Value('string'), 'granularity': Value('string'), 'created_utc': Value('timestamp[s]'), 'source': {'provider': Value('string'), 'id_list': Value('string'), 'text_root': Value('string')}, 'counts': {'candidates': Value('int64'), 'admitted': Value('int64'), 'excluded': Value('int64'), 'excluded_by_reason': {'license': Value('int64'), 'retracted': Value('int64')}, 'legacy_pdfs_excluded': Value('int64')}, 'license_summary': {'CC-BY-NC-4.0': Value('int64'), 'CC-BY-4.0': Value('int64'), 'CC-BY-3.0': Value('int64'), 'CC-BY-NC-SA-4.0': Value('int64'), 'CC-BY-NC-3.0': Value('int64'), 'CC-BY-2.0': Value('int64'), 'CC-BY-NC-SA-1.0': Value('int64')}, 'totals': {'words': Value('int64'), 'chars': Value('int64')}, 'files': List({'path': Value('string'), 'sha256': Value('string'), 'bytes': Value('int64'), 'rows': Value('int64')}), 'policy': {'redistribution': Value('string'), 'no_derivatives': Value('string'), 'retracted': Value('string')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
pmcid: string
reason: string
detail: string
license_summary: struct<CC-BY-NC-4.0: int64, CC-BY-4.0: int64, CC-BY-3.0: int64, CC-BY-NC-SA-4.0: int64, CC-BY-NC-3.0 (... 50 chars omitted)
child 0, CC-BY-NC-4.0: int64
child 1, CC-BY-4.0: int64
child 2, CC-BY-3.0: int64
child 3, CC-BY-NC-SA-4.0: int64
child 4, CC-BY-NC-3.0: int64
child 5, CC-BY-2.0: int64
child 6, CC-BY-NC-SA-1.0: int64
granularity: string
files: list<item: struct<path: string, sha256: string, bytes: int64, rows: int64>>
child 0, item: struct<path: string, sha256: string, bytes: int64, rows: int64>
child 0, path: string
child 1, sha256: string
child 2, bytes: int64
child 3, rows: int64
corpus_version: string
policy: struct<redistribution: string, no_derivatives: string, retracted: string>
child 0, redistribution: string
child 1, no_derivatives: string
child 2, retracted: string
source: struct<provider: string, id_list: string, text_root: string>
child 0, provider: string
child 1, id_list: string
child 2, text_root: string
counts: struct<candidates: int64, admitted: int64, excluded: int64, excluded_by_reason: struct<license: int6 (... 50 chars omitted)
child 0, candidates: int64
child 1, admitted: int64
child 2, excluded: int64
child 3, excluded_by_reason: struct<license: int64, retracted: int64>
child 0, license: int64
child 1, retracted: int64
child 4, legacy_pdfs_excluded: int64
created_utc: timestamp[s]
builder_version: string
artifact_id: string
totals: struct<words: int64, chars: int64>
child 0, words: int64
child 1, chars: int64
to
{'artifact_id': Value('string'), 'corpus_version': Value('string'), 'builder_version': Value('string'), 'granularity': Value('string'), 'created_utc': Value('timestamp[s]'), 'source': {'provider': Value('string'), 'id_list': Value('string'), 'text_root': Value('string')}, 'counts': {'candidates': Value('int64'), 'admitted': Value('int64'), 'excluded': Value('int64'), 'excluded_by_reason': {'license': Value('int64'), 'retracted': Value('int64')}, 'legacy_pdfs_excluded': Value('int64')}, 'license_summary': {'CC-BY-NC-4.0': Value('int64'), 'CC-BY-4.0': Value('int64'), 'CC-BY-3.0': Value('int64'), 'CC-BY-NC-SA-4.0': Value('int64'), 'CC-BY-NC-3.0': Value('int64'), 'CC-BY-2.0': Value('int64'), 'CC-BY-NC-SA-1.0': Value('int64')}, 'totals': {'words': Value('int64'), 'chars': Value('int64')}, 'files': List({'path': Value('string'), 'sha256': Value('string'), 'bytes': Value('int64'), 'rows': Value('int64')}), 'policy': {'redistribution': Value('string'), 'no_derivatives': Value('string'), 'retracted': Value('string')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
MAO Corpus V1 — source documents
A license-verified, document-level biomedical corpus assembled from PubMed Central full text. 584 documents, one row per document, every row carrying its own provenance and a positively identified redistribution license.
This dataset is the source artifact. Chunking, embeddings and sparse indexes are derived artifacts and are published separately — a corpus and an index built from it are never the same artifact.
Why this exists
It replaces a 161 MB JSON blob that labelled every document "open-access".
That string was the fallback value of a broken XPath, not a license claim. When
the licenses were actually fetched and parsed from each article's JATS
<permissions> block, 315 of 900 documents (35%) turned out not to be
redistributable, and one had been retracted. None of that was detectable in
the original data.
Contents
| File | Rows | Description |
|---|---|---|
documents.parquet |
584 | One row per admitted document: full text plus verified provenance |
corpus_manifest.json |
— | Counts, license summary, per-file SHA-256 |
exclusions.json |
316 | Every rejected candidate with an explicit reason |
provenance.json |
900 | Resolved provenance for all candidates, admitted or not |
documents.parquet schema
| Column | Type | Notes |
|---|---|---|
pmcid |
string | PMC identifier; unique, 0 duplicates |
title, journal, year, doi, pmid |
string | Fetched from NCBI efetch, not inferred |
authors |
list<string> | |
article_type |
string | JATS article type |
license_id |
string | Normalized SPDX-style identifier; never UNKNOWN |
license_url |
string | The license URL as declared by the publisher |
copyright_statement |
string | Verbatim from the article |
pmc_url |
string | Canonical source URL |
text |
string | Full document text |
sha256 |
string | Content hash of text |
word_count, char_count |
int64 |
Licensing — read this before redistributing
Licenses are per document, not per dataset. The license_id column is
authoritative for each row.
| License | Documents |
|---|---|
| CC-BY-4.0 | 460 |
| CC-BY-NC-4.0 | 102 |
| CC-BY-3.0 | 7 |
| CC-BY-NC-SA-4.0 | 7 |
| CC-BY-NC-3.0 | 6 |
| CC-BY-2.0 | 1 |
| CC-BY-NC-SA-1.0 | 1 |
| UNKNOWN | 0 |
116 documents carry a NonCommercial (NC) clause. They are recorded
separately so a consumer can partition them out; filter on license_id before
any commercial use. All licenses require attribution — the authors, title,
journal, doi and license_url columns carry what is needed to give it.
Admission policy
Fail-closed. A document is admitted only when a license is positively identified as permitting redistribution:
- Allowed: CC0, CC-BY, CC-BY-SA, CC-BY-NC, CC-BY-NC-SA.
- Refused: every ND (no-derivatives) variant — a chunked corpus is a derivative work — and anything with no identified license.
- Refused: the bare string
"open-access", which normalizes toUNKNOWN. That was the legacy fallback and is not a license. - Refused: retracted articles — not admissible as clinical evidence.
Every candidate is either admitted or recorded in exclusions.json with a
reason. Nothing is dropped silently.
Intended use and limitations
Built for retrieval research and evidence grounding. It is not a clinical decision tool and carries no medical authority.
Known limitations, measured rather than assumed:
- Document boundaries are inherited from the legacy ingestion and are imperfect. The largest admitted document is 120,783 words and is almost certainly a whole supplement ingested as one article. Re-derivation is planned.
- Topically narrow. Selection targeted stroke, Alzheimer's disease and general health; it is not a representative sample of biomedical literature.
- Retraction status reflects the fetch date recorded in
corpus_manifest.json. A paper retracted after that date will not be flagged. Re-check before relying on any document as current evidence. - Year range 2015–2026, so older foundational literature is absent.
Provenance
Fetched from NCBI E-utilities (efetch, db=pmc) against a recorded PMCID list.
License and retraction status come from each article's JATS <permissions> and
article-type metadata — actual license terms, not subset membership. The PMC OA
Web Service (oa.fcgi) is retired (HTTP 404) and is deliberately not used.
Integrity
Every file is pinned by SHA-256 and byte size in corpus_manifest.json, and the
consuming code refuses to load an artifact whose bytes disagree with its pin.
Address this dataset by an immutable commit revision, never by a branch.
- Downloads last month
- 22