world-science / scripts /science_face.py
Lilambd's picture
Upload folder using huggingface_hub
cf08d94 verified
Raw History Blame Contribute Delete
14.4 kB
# science_face.py - World Science face: every OpenAlex work counted by subfield x country x year x type.
# usage: python science_face.py BASE_DIR [--upload] (needs: pip install duckdb pyarrow requests huggingface_hub)
# -*- coding: utf-8 -*-
import sys, os, io, time, datetime, requests
BASE = sys.argv[1] if len(sys.argv) > 1 and not sys.argv[1].startswith('--') else '.'
PART = os.path.join(BASE, 'part'); O = os.path.join(BASE, 'hf_world-science')
os.makedirs(PART, exist_ok=True); os.makedirs(O, exist_ok=True)
S3 = 'https://openalex.s3.amazonaws.com/'
S = requests.Session()
if os.environ.get('REQUESTS_CA_BUNDLE'): S.verify = os.environ['REQUESTS_CA_BUNDLE']
def U(s3url): return s3url.replace('s3://openalex/', S3)
def fw(p): return p.replace('\\', '/')
class RF(io.RawIOBase):
"""HTTP range-read file, so pyarrow fetches only the needed column chunks."""
def __init__(s, u):
s.u = u; s.p = 0; s.got = 0
s.n = int(S.head(u, timeout=60).headers['Content-Length'])
def readable(s): return True
def seekable(s): return True
def seek(s, o, w=0):
s.p = o if w == 0 else (s.p + o if w == 1 else s.n + o); return s.p
def tell(s): return s.p
def read(s, k=-1):
if k < 0: k = s.n - s.p
if k == 0 or s.p >= s.n: return b''
e = min(s.n, s.p + k) - 1
for t in range(5):
try:
r = S.get(s.u, headers={'Range': f'bytes={s.p}-{e}'}, timeout=120); r.raise_for_status(); b = r.content; break
except Exception:
if t == 4: raise
time.sleep(3 * (t + 1))
s.p += len(b); s.got += len(b); return b
def readinto(s, b):
d = s.read(len(b)); b[:len(d)] = d; return len(d)
COLS = ['publication_year', 'type', 'primary_topic.subfield.id', 'authorships.list.element.countries',
'open_access.is_oa', 'is_retracted', 'created_date', 'cited_by_count']
AGG = """
with w as (select publication_year y, type ty, primary_topic.subfield.id sf,
list_distinct(flatten(list_transform(authorships, lambda x: x.countries))) cc,
open_access.is_oa oa, is_retracted rt, cited_by_count cb, strftime(created_date,'%Y-%m') cm from t),
e as (select *, case when len(list_filter(cc, lambda x: x is not null))=0 then ['XX'] else list_filter(cc, lambda x: x is not null) end c2 from w),
u as (select y,ty,sf,oa,rt,cb,cm,unnest(c2) c, 1.0/len(c2) fr from e)
select coalesce(y,0) y,
case when ty in ('article','review','preprint','dataset','dissertation','letter','editorial') then ty
when ty in ('book','book-chapter','monograph','reference-entry') then 'book'
when ty in ('paratext','erratum','retraction','supplementary-materials','peer-review') then 'paratext' else 'other' end tc,
coalesce(replace(sf,'https://openalex.org/subfields/',''),'none') sf, c,
count(*) n, sum(fr) nf, sum(case when oa then 1 else 0 end) n_oa, sum(case when rt then 1 else 0 end) n_retracted,
sum(coalesce(cb,0)) cites, cm
from u group by all"""
def one(a):
import pyarrow.parquet as pq, duckdb
i, url = a
o = os.path.join(PART, f'{i:05d}.parquet')
if os.path.exists(o): return i, 0
for t in range(4):
try:
f = RF(U(url)); tb = pq.ParquetFile(f).read(columns=COLS); break
except Exception as ex:
if t == 3: return i, 'ERR ' + str(ex)[:200]
time.sleep(5 * (t + 1))
c = duckdb.connect(); c.register('t', tb)
pq.write_table(c.sql(AGG).to_arrow_table(), o + '.tmp'); os.replace(o + '.tmp', o)
return i, f.got
def lookups(m):
import pyarrow as pa, pyarrow.parquet as pq, duckdb
for ent, fn, q in [
('subfields', 'subfields.csv', "select replace(id,'https://openalex.org/subfields/','') subfield_id, display_name subfield, replace(field.id,'https://openalex.org/fields/','') field_id, field.display_name field, replace(domain.id,'https://openalex.org/domains/','') domain_id, domain.display_name AS \"domain\" from t order by 1"),
('countries', 'countries.csv', "select country_code, display_name country, continent.display_name continent, is_global_south, alpha_3 from t order by 1")]:
e = [x for x in m['entities'] if x['entity'] == ent][0]
t = pa.concat_tables([pq.read_table(RF(U(f['url']))) for f in e['files']], promote_options='default')
c = duckdb.connect(); c.register('t', t)
c.sql(f"copy ({q}) to '{fw(O)}/{fn}' (header)")
def build(SNAP):
import duckdb
P = fw(PART) + '/*.parquet'; OO = fw(O)
FET = datetime.datetime.now(datetime.timezone.utc).strftime('%Y-%m-%dT%H:%MZ')
SRC = 'OpenAlex works snapshot ' + SNAP + ' (CC0), s3://openalex/data/parquet/works/'
c = duckdb.connect()
c.sql(f"create table a as select y,tc,sf,c,sum(n) n,sum(nf) nf,sum(n_oa) n_oa,sum(n_retracted) n_rt,sum(cites) cites from '{P}' group by all")
c.sql(f"create table m as select cm,tc,sf,c,sum(n) n,sum(nf) nf from '{P}' where cm>='2025-11' group by all")
c.sql(f"create table sfl as select * from read_csv('{OO}/subfields.csv')")
c.sql(f"create table cl as select * from read_csv('{OO}/countries.csv')")
tail = f", '{SRC}' AS source, '{SNAP}' AS snapshot_date, '{FET}' AS fetched_at"
J = "left join sfl on try_cast(a.sf as int)=sfl.subfield_id left join cl on a.c=cl.country_code"
PAP = "('article','review','book','letter','editorial','dissertation')"
def out(name, q, csv=True):
c.sql(f"copy ({q}) to '{OO}/{name}.parquet' (format parquet)")
if csv: c.sql(f"copy ({q}) to '{OO}/{name}.csv' (header)")
print(name, c.sql(f"select count(*) from '{OO}/{name}.parquet'").fetchall()[0], flush=True)
out('works_by_subfield_country_year_type', f"""select a.y AS "year", a.tc AS work_type, a.sf AS subfield_id, sfl.subfield, sfl.field, sfl.domain, a.c AS country_code, cl.country,
a.n AS works, round(a.nf,3) AS works_fractional, a.n_oa AS works_oa, a.n_rt AS retracted, a.cites AS citations{tail}
from a {J} where a.y between 1900 and 2026 order by 1,2,3,7""", csv=False)
out('papers_by_field_year', f"""select a.y AS "year", sfl.field_id, sfl.field, sfl.domain,
round(sum(case when tc in {PAP} then nf end)) AS papers, round(sum(case when tc='preprint' then nf end)) AS preprints,
round(sum(case when tc='dataset' then nf end)) AS datasets, round(sum(case when tc not in {PAP} and tc not in ('preprint','dataset') then nf end)) AS other_records,
round(log10(nullif(sum(case when tc in {PAP} then nf end),0)),2) AS log10_papers{tail}
from a {J} where a.y between 1900 and 2026 group by all order by 1,2""")
out('papers_by_country_year', f"""select a.y AS "year", a.c AS country_code, cl.country, cl.continent, cl.is_global_south,
sum(case when tc in {PAP} then n end) AS papers, round(sum(case when tc in {PAP} then nf end),1) AS papers_fractional,
sum(case when tc='preprint' then n end) AS preprints, sum(case when tc='dataset' then n end) AS datasets,
sum(case when tc in {PAP} then n_oa end) AS papers_oa, sum(case when tc in {PAP} then cites end) AS citations_of_papers{tail}
from a {J} where a.y between 1900 and 2026 group by all order by 1,2""")
out('new_records_by_month', f"""select a.cm AS month_added, a.tc AS work_type, sfl.field, a.c AS country_code, cl.country, sum(a.n) AS works, round(sum(a.nf),2) AS works_fractional{tail}
from m a {J} group by all order by 1,2,3,4""")
out('rising_subfield_country_2025', f"""with b as (select sf,c, sum(case when y=2025 and tc in {PAP} then n end) n25, sum(case when y between 2020 and 2024 and tc in {PAP} then n end)/5.0 base from a group by all)
select b.sf AS subfield_id, sfl.subfield, sfl.field, b.c AS country_code, cl.country, b.n25 AS papers_2025, round(b.base,1) AS mean_papers_2020_2024,
round(100.0*b.n25/nullif(b.base,0)) AS index_2025_vs_base100, round(log10(b.n25),2) AS log10_papers_2025{tail}
from b left join sfl on try_cast(b.sf as int)=sfl.subfield_id left join cl on b.c=cl.country_code
where b.n25>=100 and b.base>=20 and b.c<>'XX' and b.sf<>'none' order by index_2025_vs_base100 desc""")
README = r'''---
license: cc0-1.0
pretty_name: World Science (papers by field x country x year, from OpenAlex)
language:
- en
- ja
tags:
- science
- bibliometrics
- openalex
- research
- countries
- monthly
size_categories:
- 1M<n<10M
configs:
- config_name: papers_by_field_year
data_files: papers_by_field_year.csv
- config_name: papers_by_country_year
data_files: papers_by_country_year.csv
- config_name: rising_subfield_country_2025
data_files: rising_subfield_country_2025.csv
- config_name: new_records_by_month
data_files: new_records_by_month.parquet
- config_name: works_by_subfield_country_year_type
data_files: works_by_subfield_country_year_type.parquet
- config_name: subfields
data_files: subfields.csv
- config_name: countries
data_files: countries.csv
---
# World Science — every OpenAlex work counted by field x country x year x type
**All 464,765,485 works** in the OpenAlex snapshot of 2026-10-14, counted once each (the fractional totals add back to the snapshot record count). No sampling, no model, no API ranking: the whole snapshot was read column by column from the public OpenAlex S3 bucket and grouped.
Part of the [World Signals](https://huggingface.co/datasets/Lilambd/world-signals) family: one flat table per way of looking at the world. This one is the science face: where research is written, in which of 252 subfields, and what is rising.
Every row carries `source`, `snapshot_date` and `fetched_at`.
Quick start:
```python
from datasets import load_dataset
fy = load_dataset("Lilambd/world-science", "papers_by_field_year", split="train")
rising = load_dataset("Lilambd/world-science", "rising_subfield_country_2025", split="train")
```
| file | rows | what |
|---|---|---|
| `works_by_subfield_country_year_type.parquet` | 3.43M | year (1900-2026) x work type x subfield (252 + none) x author country. `works` = full counting (a work with authors in JP and US counts once for each), `works_fractional` = 1/k per country so totals do not double count. Also open-access count, retracted count, citations. |
| `papers_by_field_year.csv` | 3,429 | world total per field and year: papers, preprints, datasets, other records, log10 of papers |
| `papers_by_country_year.csv` | 18,007 | per country and year: papers (full and fractional), preprints, datasets, open-access papers, citations |
| `rising_subfield_country_2025.csv` | 8,716 | subfield x country cells where 2025 papers are compared with the 2020-2024 mean (index, base = 100). Only cells with at least 100 papers in 2025 and a mean of at least 20 |
| `new_records_by_month.parquet` | 204k | records that entered OpenAlex each month from 2025-11 to 2026-10, by type, field and country |
| `subfields.csv`, `countries.csv` | | lookups (subfield -> field -> domain; country -> continent, Global South flag) |
"papers" = article, review, book / chapter, letter, editorial, dissertation. Preprints and datasets are counted separately on purpose.
## Read before using (measured, not assumed)
- **Datasets swamp the totals from 2024.** OpenAlex type `dataset` was 24.9M records for publication year 2025, against 8.4M articles. About 18.1M of those 2025 datasets carry a Japanese affiliation, almost all in Nuclear and High Energy Physics and Aerospace Engineering. Count papers, not works, when comparing countries.
- **No country for many works.** `country_code = XX` means no author affiliation country is recorded. For 2025 that is 4.3M fractional papers, more than any single country.
- **Recent years are still filling.** 2026 is a partial year, and OpenAlex keeps adding records for past years.
- **Month added starts 2025-11.** OpenAlex rebuilt its index in 2025-10 and gave about 356M records that creation month, so earlier months are not a real flow and are left out.
- About 17.7M works with no publication year, or a year before 1900 or after 2026, are left out of the year tables.
- Country = country of the author's affiliation as OpenAlex records it, not nationality.
## Refresh
Rebuilt from each new OpenAlex snapshot (about monthly). The script that made it is `scripts/science_face.py` (it reads only the needed columns over HTTP range requests, about 2.4MB per 750MB file, 4.6GB for the whole snapshot, then builds the tables).
## Attribution
Data from [OpenAlex](https://openalex.org) (Priem, Piwowar and Orr, 2022), released CC0. Grouping and tables by [Lilambd](https://huggingface.co/Lilambd), also CC0. Questions or a slice you need: hello@lilambd.com
---
日本語: OpenAlex の論文など 4.65億件 (2026-10-14 時点) を全部、分野252 × 国 × 年 × 種類 で数えた台帳。2024年からは「データセット」の登録が論文より多く、2025年は日本の所属で約1,810万件 (ほぼ物理と航空宇宙) あるため、国を比べる時は papers を使う。
'''
if __name__ == '__main__':
from multiprocessing import Pool
m = S.get(S3 + 'data/parquet/manifest.json', timeout=120).json(); SNAP = m['date']
works = [e for e in m['entities'] if e['entity'] == 'works'][0]
fs = [(i, f['url']) for i, f in enumerate(works['files'])]
print('snapshot', SNAP, 'files', len(fs), 'records', works['record_count'], flush=True)
t0 = time.time(); n = 0; bad = 0
with Pool(int(os.environ.get('NP', '8'))) as P:
for i, g in P.imap_unordered(one, fs):
n += 1
if isinstance(g, str): bad += 1; print(i, g, flush=True)
if n % 100 == 0: print(n, '/', len(fs), round(time.time() - t0), 's', flush=True)
if bad: print('FAILED files:', bad, '- run again to retry only those'); sys.exit(1)
lookups(m); build(SNAP)
open(os.path.join(O, 'README.md'), 'w', encoding='utf-8', newline='\n').write(README)
os.makedirs(os.path.join(O, 'scripts'), exist_ok=True)
import shutil; shutil.copy(os.path.abspath(__file__), os.path.join(O, 'scripts', 'science_face.py'))
if '--upload' in sys.argv:
from huggingface_hub import HfApi
a = HfApi(); r = 'Lilambd/world-science'
a.create_repo(r, repo_type='dataset', exist_ok=True)
a.upload_folder(repo_id=r, repo_type='dataset', folder_path=O, ignore_patterns=['part/*'])
print('OK https://huggingface.co/datasets/' + r)
print('Done')