empire-eval / scripts /ds_hazard.py
Lilambd's picture
Publish Lilambd Empire Eval
7f9a652 verified
Raw History Blame Contribute Delete
15 kB
from evallib import fmt
from common import long_features
DS = 'Lilambd/world-hazard'
REVISION = 'hf-main@2026-11-03 (data as of 2026-11-01)'
LICENSE = 'CC BY 4.0'
STR = {'data/admin1.csv': ['a1', 'iso2', 'iso_3166_2', 'hasc'], 'data/station_warming.csv': ['station', 'fips2']}
KEYS = {
'data/admin1.csv': ['a1', 'name', 'country', 'iso_3166_2'],
'data/admin1_quake.csv': ['a1', 'name', 'country'],
'data/admin1_warming.csv': ['a1', 'name', 'country'],
'data/country_warming.csv': ['country'],
'data/country_quake.csv': ['method', 'country'],
'data/country_event.csv': ['kind', 'country'],
'data/decade_quake.csv': ['decade'],
'data/band_warming.csv': ['band'],
'data/station_warming.csv': ['station', 'name'],
'data/events.csv': ['kind', 'id', 'year', 'country'],
'data/joint_exposure.csv': ['a1', 'name', 'country'],
'data/features_country.csv': ['iso3', 'feature'],
'data/year_anomaly.csv': ['year', 'band'],
}
SH = 'World Hazard Ledger'
def build(B):
s = B.store
q = s['data/admin1_quake.csv']
# unique (name,country)
dup = q.duplicated(['name', 'country'], keep=False)
qu = q[~dup]
for _, r in qu.loc[B.sample(qu.index, 45)].iterrows():
place = f"{r['name']} ({r['country']})"
c = B.rng.choice(['m6', 'm7', 'm6_per_1000km2', 'm6_in'])
text = {
'm6': f'In the {SH}, how many M6+ earthquakes from 1900 to 2026 are assigned to {place}?',
'm7': f'In the {SH}, how many M7+ earthquakes from 1900 to 2026 are assigned to {place}?',
'm6_per_1000km2': f'In the {SH}, how many M6+ earthquakes (1900-2026) per 1,000 km² does {place} have?',
'm6_in': f'In the {SH}, how many of the M6+ earthquakes assigned to {place} had their epicentre inside the region (column m6_in)?',
}[c]
B.add('lookup', text, {'op': 'lookup', 'file': 'data/admin1_quake.csv', 'where': [['name', '==', r['name']], ['country', '==', r['country']]], 'col': c},
tol={'abs': 0} if c != 'm6_per_1000km2' else {'rel': 0.01}, skills=['table_lookup'])
cs = q.groupby('country').size()
for c in B.sample(cs[cs >= 4].index, 22):
col = B.rng.choice(['m6', 'm7', 'm6_per_1000km2'])
word = {'m6': 'the most M6+ earthquakes', 'm7': 'the most M7+ earthquakes', 'm6_per_1000km2': 'the highest density of M6+ earthquakes per 1,000 km²'}[col]
B.add('rank', f'In the {SH}, which admin-1 region of {c} has {word} since 1900?',
{'op': 'argmax', 'file': 'data/admin1_quake.csv', 'where': [['country', '==', c]], 'col': col, 'return': 'name'},
difficulty=2, skills=['ranking', 'group_filter'])
B.add('rank', f'In the {SH}, list the 5 admin-1 regions with the most M7+ earthquakes since 1900, most first.',
{'op': 'topk', 'k': 5, 'file': 'data/admin1_quake.csv', 'col': 'm7', 'return': 'name'}, difficulty=3, skills=['ranking', 'full_scan'], answer_format='JSON list of region names')
for _ in range(18):
a, b = qu.sample(2, random_state=B.rng.randint(0, 1 << 30)).to_dict('records')
B.add('compare', f"In the {SH}, which region had more M6+ earthquakes since 1900: {a['name']} ({a['country']}) or {b['name']} ({b['country']})?",
{'op': 'compare', 'labels': [a['name'], b['name']], 'parts': [
{'op': 'lookup', 'file': 'data/admin1_quake.csv', 'where': [['a1', '==', a['a1']]], 'col': 'm6'},
{'op': 'lookup', 'file': 'data/admin1_quake.csv', 'where': [['a1', '==', b['a1']]], 'col': 'm6'}]},
difficulty=2, skills=['comparison'])
for c in B.sample(cs[cs >= 3].index, 12):
B.add('aggregate', f'In the {SH}, what is the total number of M7+ earthquakes summed over all admin-1 regions of {c} listed in admin1_quake?',
{'op': 'sum', 'file': 'data/admin1_quake.csv', 'where': [['country', '==', c]], 'col': 'm7'}, tol={'abs': 0}, difficulty=2, skills=['aggregation'])
B.add('count', f'In the {SH}, how many admin-1 regions of {c} have at least one M7+ earthquake since 1900?',
{'op': 'count', 'file': 'data/admin1_quake.csv', 'where': [['country', '==', c], ['m7', '>=', 1]]}, difficulty=2, skills=['filter', 'count'])
w = s['data/admin1_warming.csv']
wu = w[~w.duplicated(['name', 'country'], keep=False)]
for _, r in wu.loc[B.sample(wu.index, 35)].iterrows():
col = B.rng.choice(['mean', 'mean', 'n'])
t = (f"In the {SH}, by how many °C did the weather stations of {r['name']} ({r['country']}) warm on average (1991-2020 normal minus 1961-1990 normal)?"
if col == 'mean' else f"In the {SH}, how many weather stations with both normals are counted for {r['name']} ({r['country']})?")
B.add('lookup', t, {'op': 'lookup', 'file': 'data/admin1_warming.csv', 'where': [['a1', '==', r['a1']]], 'col': col},
unit='°C' if col == 'mean' else None, tol={'abs': 0.002} if col == 'mean' else {'abs': 0}, skills=['table_lookup'])
wc = w.groupby('country').size()
for c in B.sample(wc[wc >= 4].index, 15):
B.add('rank', f'In the {SH}, which admin-1 region of {c} shows the largest mean station warming?',
{'op': 'argmax', 'file': 'data/admin1_warming.csv', 'where': [['country', '==', c]], 'col': 'mean', 'return': 'name'}, difficulty=2, skills=['ranking', 'group_filter'])
# absent warming: regions in admin1 with no station record
a1 = s['data/admin1.csv']
have = set(w['a1'])
miss = a1[~a1['a1'].isin(have) & ~a1.duplicated(['name', 'country'], keep=False)]
for _, r in miss.loc[B.sample(miss.index, 14)].iterrows():
B.add('unanswerable', f"In the {SH}, what is the mean station warming (°C) recorded for {r['name']} ({r['country']})? If the ledger has no station record for it, answer null.",
{'op': 'absent', 'file': 'data/admin1_warming.csv', 'where': [['a1', '==', r['a1']]]}, difficulty=2, skills=['abstention'],
note='The README states a region missing from admin1_warming has no station record, which is not the same as no warming.')
cw = s['data/country_warming.csv']
for _, r in cw.loc[B.sample(cw.index, 15)].iterrows():
B.add('lookup', f"In the {SH}, what is the mean station warming of {r['country']} in °C?",
{'op': 'lookup', 'file': 'data/country_warming.csv', 'where': [['country', '==', r['country']]], 'col': 'mean'}, unit='°C', tol={'abs': 0.002}, skills=['table_lookup'])
for _ in range(10):
a, b = cw.sample(2, random_state=B.rng.randint(0, 1 << 30))['country']
B.add('compare', f'In the {SH}, which country warmed more on average (station mean): {a} or {b}?',
{'op': 'compare', 'labels': [a, b], 'parts': [
{'op': 'lookup', 'file': 'data/country_warming.csv', 'where': [['country', '==', a]], 'col': 'mean'},
{'op': 'lookup', 'file': 'data/country_warming.csv', 'where': [['country', '==', b]], 'col': 'mean'}]}, difficulty=2, skills=['comparison'])
B.add('rank', f'In the {SH}, which country has the highest mean station warming among countries with at least 10 stations?',
{'op': 'argmax', 'file': 'data/country_warming.csv', 'where': [['n', '>=', 10]], 'col': 'mean', 'return': 'country'}, difficulty=3, skills=['ranking', 'filter'])
ce = s['data/country_event.csv']
kinds = sorted(ce['kind'].unique())
for _, r in ce.loc[B.sample(ce.index, 25)].iterrows():
col = B.rng.choice(['n', 'deaths'])
kind = r['kind'].replace('sig_', 'significant ')
t = (f"In the {SH}, how many {kind} events are recorded for {r['country']}?" if col == 'n' else f"In the {SH}, how many deaths from {kind} events are recorded for {r['country']}?")
B.add('lookup', t, {'op': 'lookup', 'file': 'data/country_event.csv', 'where': [['kind', '==', r['kind']], ['country', '==', r['country']]], 'col': col}, tol={'abs': 0}, skills=['table_lookup'])
for k in kinds:
kind = k.replace('sig_', 'significant ')
B.add('rank', f'In the {SH}, which country has the most recorded deaths from {kind} events?',
{'op': 'argmax', 'file': 'data/country_event.csv', 'where': [['kind', '==', k]], 'col': 'deaths', 'return': 'country'}, difficulty=2, skills=['ranking'])
B.add('rank', f'In the {SH}, list the 3 countries with the most {kind} events, most first.',
{'op': 'topk', 'k': 3, 'file': 'data/country_event.csv', 'where': [['kind', '==', k]], 'col': 'n', 'return': 'country'}, difficulty=2, skills=['ranking'], answer_format='JSON list of country names as written in the ledger')
B.add('aggregate', f'In the {SH}, what is the total number of deaths from {kind} events summed over all countries?',
{'op': 'sum', 'file': 'data/country_event.csv', 'where': [['kind', '==', k]], 'col': 'deaths'}, tol={'abs': 0}, difficulty=2, skills=['aggregation'])
for _, r in s['data/decade_quake.csv'].iterrows():
col = B.rng.choice(['m6', 'm7', 'm8'])
B.add('lookup', f"In the {SH}, how many {col.upper()}+ earthquakes are counted in the decade starting {r['decade']}?",
{'op': 'lookup', 'file': 'data/decade_quake.csv', 'where': [['decade', '==', int(r['decade'])]], 'col': col}, tol={'abs': 0}, skills=['table_lookup'])
B.add('rank', f'In the {SH}, which decade has the most M8+ earthquakes?', {'op': 'argmax', 'file': 'data/decade_quake.csv', 'col': 'm8', 'return': 'decade'}, difficulty=2, skills=['ranking'])
for _, r in s['data/band_warming.csv'].iterrows():
B.add('lookup', f"In the {SH}, what is the mean station warming (°C) in the latitude band {r['band']}?",
{'op': 'lookup', 'file': 'data/band_warming.csv', 'where': [['band', '==', r['band']]], 'col': 'mean'}, unit='°C', tol={'abs': 0.002}, skills=['table_lookup'])
st = s['data/station_warming.csv']
stu = st[~st.duplicated(['name'], keep=False)]
for _, r in stu.loc[B.sample(stu.index, 20)].iterrows():
B.add('lookup', f"In the {SH}, how much did GHCN-M station {r['station']} ({r['name']}) warm between the 1961-1990 and 1991-2020 normals, in °C?",
{'op': 'lookup', 'file': 'data/station_warming.csv', 'where': [['station', '==', r['station']]], 'col': 'delta'}, unit='°C', tol={'abs': 0.002}, skills=['table_lookup', 'id_lookup'])
fc = st.groupby('fips2').size()
for f in B.sample(fc[fc >= 15].index, 10):
B.add('count', f'In the {SH} station table, how many stations with FIPS country code {f} cooled (delta below 0)?',
{'op': 'count', 'file': 'data/station_warming.csv', 'where': [['fips2', '==', f], ['delta', '<', 0]]}, difficulty=2, skills=['filter', 'count'])
B.add('aggregate', f'In the {SH} station table, what is the mean warming (delta, °C) over all stations with FIPS country code {f}?',
{'op': 'mean', 'file': 'data/station_warming.csv', 'where': [['fips2', '==', f]], 'col': 'delta'}, unit='°C', tol={'abs': 0.005}, difficulty=2, skills=['aggregation'])
B.add('count', f'In the {SH}, how many weather stations got cooler (delta below 0) between the two normals?',
{'op': 'count', 'file': 'data/station_warming.csv', 'where': [['delta', '<', 0]]}, difficulty=2, skills=['filter', 'count', 'full_scan'])
ev = s['data/events.csv']
eq = ev[ev.kind == 'sig_earthquake']
ec = eq.groupby('country').size()
for c in B.sample(ec[ec >= 10].index, 14):
y = B.rng.choice([1900, 1950, 2000])
B.add('count', f'In the {SH} events table, how many significant earthquakes are recorded for {c} from {y} onward?',
{'op': 'count', 'file': 'data/events.csv', 'where': [['kind', '==', 'sig_earthquake'], ['country', '==', c], ['year', '>=', y]]}, difficulty=2, skills=['filter', 'count', 'temporal'])
B.add('rank', f'In the {SH} events table, in which year did the deadliest significant earthquake recorded for {c} occur (by deaths)?',
{'op': 'argmax', 'file': 'data/events.csv', 'where': [['kind', '==', 'sig_earthquake'], ['country', '==', c]], 'col': 'deaths', 'return': 'year'}, difficulty=3, skills=['ranking', 'group_filter'])
je = s['data/joint_exposure.csv']
B.add('count', f'In the {SH}, how many regions are in the top quarter for both M6 density and warming (joint_exposure)?', {'op': 'count', 'file': 'data/joint_exposure.csv'}, skills=['count'])
B.add('rank', f'In the {SH}, which joint-exposure region (top quarter for both shaking and warming) has the largest population?',
{'op': 'argmax', 'file': 'data/joint_exposure.csv', 'col': 'population', 'return': 'name'}, difficulty=2, skills=['ranking'])
B.add('rank', f'In the {SH}, which joint-exposure region warmed the most?', {'op': 'argmax', 'file': 'data/joint_exposure.csv', 'col': 'warming', 'return': 'name'}, difficulty=2, skills=['ranking'])
jc = je.groupby('country').size()
for c in jc[jc >= 2].index:
B.add('count', f'In the {SH}, how many joint-exposure regions are in {c}?', {'op': 'count', 'file': 'data/joint_exposure.csv', 'where': [['country', '==', c]]}, skills=['filter', 'count'])
# multi-hop: Japanese name / ISO 3166-2 -> a1 -> quake / warming
aq = a1[a1['a1'].isin(set(q['a1'])) & a1['name_ja'].notna() & ~a1.duplicated(['name_ja'], keep=False)]
for _, r in aq.loc[B.sample(aq.index, 14)].iterrows():
B.add('multi_hop', f"In the {SH}, the region whose Japanese name is 「{r['name_ja']}」 has how many M6+ earthquakes since 1900?",
{'op': 'chain', 'steps': [
{'op': 'lookup', 'file': 'data/admin1.csv', 'where': [['name_ja', '==', r['name_ja']]], 'col': 'a1', 'as': 'a1'},
{'op': 'lookup', 'file': 'data/admin1_quake.csv', 'where': [['a1', '==', '$a1']], 'col': 'm6'}]},
tol={'abs': 0}, difficulty=3, skills=['join', 'multilingual_entity'])
aw = a1[a1['a1'].isin(set(w['a1'])) & a1['iso_3166_2'].notna() & ~a1.duplicated(['iso_3166_2'], keep=False) & ~a1['iso_3166_2'].astype(str).str.contains('~')]
for _, r in aw.loc[B.sample(aw.index, 14)].iterrows():
B.add('multi_hop', f"In the {SH}, what is the mean station warming (°C) of the region with ISO 3166-2 code {r['iso_3166_2']}?",
{'op': 'chain', 'steps': [
{'op': 'lookup', 'file': 'data/admin1.csv', 'where': [['iso_3166_2', '==', r['iso_3166_2']]], 'col': 'a1', 'as': 'a1'},
{'op': 'lookup', 'file': 'data/admin1_warming.csv', 'where': [['a1', '==', '$a1']], 'col': 'mean'}]},
unit='°C', tol={'abs': 0.002}, difficulty=3, skills=['join', 'code_resolution'])
long_features(B, 'data/features_country.csv', n_lookup=30, n_rank=4, n_cmp=8, n_count=4, n_absent=8, ds_short=SH)