empire-eval / scripts /ds_science.py
Lilambd's picture
Publish Lilambd Empire Eval
7f9a652 verified
Raw History Blame Contribute Delete
9.25 kB
DS = 'Lilambd/world-science'
REVISION = 'hf-main@2026-11-03 (OpenAlex snapshot 2026-10-14)'
LICENSE = 'CC0 1.0'
SH = 'World Science dataset (OpenAlex counts)'
STR = {'papers_by_country_year.csv': ['country_code'], 'countries.csv': ['country_code'], 'rising_subfield_country_2025.csv': ['country_code']}
KEYS = {'papers_by_country_year.csv': ['year', 'country_code', 'country'], 'papers_by_field_year.csv': ['year', 'field'],
'rising_subfield_country_2025.csv': ['subfield_id', 'country_code'], 'subfields.csv': ['subfield_id', 'subfield'], 'countries.csv': ['country_code', 'country']}
COL = {'papers': 'how many papers', 'preprints': 'how many preprints', 'datasets': 'how many dataset records', 'papers_oa': 'how many open-access papers', 'citations_of_papers': 'how many citations to its papers'}
def build(B):
s = B.store
p = s['papers_by_country_year.csv']
pr = p[(p.year >= 1990) & p.papers.notna() & (p.papers > 50)]
for r in pr.loc[B.sample(pr.index, 70)].itertuples():
col = B.rng.choice(['papers', 'papers', 'papers_oa', 'preprints', 'citations_of_papers'])
if getattr(r, col) != getattr(r, col): # NaN
continue
B.add('lookup', f'In the {SH}, {COL[col]} are counted for {r.country} in {r.year}?',
{'op': 'lookup', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', r.country_code], ['year', '==', int(r.year)]], 'col': col}, tol={'rel': 0.001}, skills=['table_lookup', 'temporal'])
for _ in range(25):
yr = B.rng.randint(1950, 2025)
cont = B.rng.choice(sorted(p.continent.dropna().unique()))
B.add('rank', f'In the {SH}, which country in {cont} had the most papers in {yr}?',
{'op': 'argmax', 'file': 'papers_by_country_year.csv', 'where': [['year', '==', yr], ['continent', '==', cont]], 'col': 'papers', 'return': 'country'}, difficulty=2, skills=['ranking', 'group_filter'])
for yr in B.sample(range(1980, 2026), 10):
B.add('rank', f'In the {SH}, list the top 3 countries by number of papers in {yr}, most first.',
{'op': 'topk', 'k': 3, 'file': 'papers_by_country_year.csv', 'where': [['year', '==', yr]], 'col': 'papers', 'return': 'country'}, difficulty=2, skills=['ranking'], answer_format='JSON list of country names')
big = sorted(p[(p.year == 2024) & (p.papers > 5000)].country_code)
for cc in B.sample(big, 22):
y1 = B.rng.randint(2000, 2014)
y2 = B.rng.randint(2015, 2024)
nm = p[p.country_code == cc].country.iloc[0]
B.add('temporal', f'In the {SH}, by what factor did the paper count of {nm} grow from {y1} to {y2} (papers {y2} divided by papers {y1})?',
{'op': 'ratio', 'parts': [
{'op': 'lookup', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', cc], ['year', '==', y2]], 'col': 'papers'},
{'op': 'lookup', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', cc], ['year', '==', y1]], 'col': 'papers'}]}, tol={'rel': 0.01}, difficulty=2, skills=['arithmetic', 'temporal'])
B.add('derived', f'In the {SH}, what share of the papers counted for {nm} in {y2} were open access (papers_oa / papers)?',
{'op': 'ratio', 'parts': [
{'op': 'lookup', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', cc], ['year', '==', y2]], 'col': 'papers_oa'},
{'op': 'lookup', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', cc], ['year', '==', y2]], 'col': 'papers'}]}, tol={'rel': 0.01}, difficulty=2, skills=['arithmetic'])
for _ in range(18):
a, b = B.sample(big, 2)
yr = B.rng.randint(2000, 2024)
na, nb = (p[p.country_code == x].country.iloc[0] for x in (a, b))
B.add('compare', f'In the {SH}, which country had more papers in {yr}: {na} or {nb}?',
{'op': 'compare', 'labels': [na, nb], 'parts': [
{'op': 'lookup', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', a], ['year', '==', yr]], 'col': 'papers'},
{'op': 'lookup', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', b], ['year', '==', yr]], 'col': 'papers'}]}, difficulty=2, skills=['comparison'])
for cc in B.sample(big, 12):
nm = p[p.country_code == cc].country.iloc[0]
B.add('aggregate', f'In the {SH}, how many papers in total are counted for {nm} over the years 2020 to 2024?',
{'op': 'sum', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', cc], ['year', '>=', 2020], ['year', '<=', 2024]], 'col': 'papers'}, tol={'rel': 0.001}, difficulty=2, skills=['aggregation', 'temporal'])
for yr in B.sample(range(2000, 2025), 8):
B.add('count', f'In the {SH}, how many countries had more than 10,000 papers in {yr}?',
{'op': 'count', 'file': 'papers_by_country_year.csv', 'where': [['year', '==', yr], ['papers', '>', 10000]]}, difficulty=2, skills=['filter', 'count'])
for cc in B.sample(big, 10):
nm = p[p.country_code == cc].country.iloc[0]
yr = B.rng.choice([2027, 2028, 2030])
B.add('unanswerable', f'In the {SH}, how many papers are counted for {nm} in {yr}? If the dataset has no such year, answer null.',
{'op': 'absent', 'file': 'papers_by_country_year.csv', 'where': [['country_code', '==', cc], ['year', '==', yr]]}, skills=['abstention', 'temporal'])
f = s['papers_by_field_year.csv']
for r in f[f.year >= 1950].pipe(lambda d: d.loc[B.sample(d.index, 30)]).itertuples():
B.add('lookup', f'In the {SH}, how many papers in the field "{r.field}" are counted for {r.year}?',
{'op': 'lookup', 'file': 'papers_by_field_year.csv', 'where': [['field', '==', r.field], ['year', '==', int(r.year)]], 'col': 'papers'}, tol={'rel': 0.001}, skills=['table_lookup'])
for yr in B.sample(range(1950, 2026), 12):
B.add('rank', f'In the {SH}, which field had the most papers in {yr}?',
{'op': 'argmax', 'file': 'papers_by_field_year.csv', 'where': [['year', '==', yr]], 'col': 'papers', 'return': 'field'}, difficulty=2, skills=['ranking'])
dom = B.rng.choice(sorted(f.domain.dropna().unique()))
B.add('aggregate', f'In the {SH}, how many papers in total are counted for all fields of the domain "{dom}" in {yr}?',
{'op': 'sum', 'file': 'papers_by_field_year.csv', 'where': [['year', '==', yr], ['domain', '==', dom]], 'col': 'papers'}, tol={'rel': 0.001}, difficulty=2, skills=['aggregation'])
rs = s['rising_subfield_country_2025.csv']
for r in rs.loc[B.sample(rs.index, 25)].itertuples():
col = B.rng.choice(['papers_2025', 'index_2025_vs_base100'])
t = (f'In the {SH} rising-subfield table, how many 2025 papers in "{r.subfield}" are counted for {r.country}?' if col == 'papers_2025'
else f'In the {SH} rising-subfield table, what is the 2025 index (2020-2024 mean = 100) of "{r.subfield}" in {r.country}?')
B.add('lookup', t, {'op': 'lookup', 'file': 'rising_subfield_country_2025.csv', 'where': [['subfield_id', '==', int(r.subfield_id)], ['country_code', '==', r.country_code]], 'col': col}, tol={'rel': 0.005}, skills=['table_lookup'])
cc_r = rs.groupby('country').size()
for c in B.sample(cc_r[cc_r >= 10].index, 15):
B.add('rank', f'In the {SH} rising-subfield table, which subfield has the highest 2025 index (vs 2020-2024 = 100) for {c}?',
{'op': 'argmax', 'file': 'rising_subfield_country_2025.csv', 'where': [['country', '==', c]], 'col': 'index_2025_vs_base100', 'return': 'subfield'}, difficulty=2, skills=['ranking', 'group_filter'])
sf = s['subfields.csv']
for r in sf.loc[B.sample(sf.index, 15)].itertuples():
B.add('lookup', f'In the {SH} taxonomy, which field does the subfield "{r.subfield}" belong to?',
{'op': 'lookup', 'file': 'subfields.csv', 'where': [['subfield_id', '==', int(r.subfield_id)]], 'col': 'field'}, skills=['taxonomy'])
for fld in B.sample(sorted(sf.field.unique()), 10):
B.add('count', f'In the {SH} taxonomy, how many subfields does the field "{fld}" have?',
{'op': 'count', 'file': 'subfields.csv', 'where': [['field', '==', fld]]}, skills=['taxonomy', 'count'])
co = s['countries.csv']
for r in co.loc[B.sample(co.index, 8)].itertuples():
B.add('lookup', f'In the {SH} country list, is {r.country} flagged as Global South (true/false)?',
{'op': 'lookup', 'file': 'countries.csv', 'where': [['country_code', '==', r.country_code]], 'col': 'is_global_south'}, skills=['table_lookup'])
for r in sf.loc[B.sample(sf.index, 10)].itertuples():
B.add('multi_hop', f'In the {SH}, how many 2025 papers in the subfield with id {r.subfield_id} are counted for the country with the most such papers? (Use rising_subfield_country_2025.)',
{'op': 'max', 'file': 'rising_subfield_country_2025.csv', 'where': [['subfield_id', '==', int(r.subfield_id)]], 'col': 'papers_2025'}, tol={'abs': 0}, difficulty=3, skills=['id_lookup', 'aggregation'])