Datasets:
Download scripts/ds_ground.py from Lilambd/empire-eval: direct link, hf CLI and curl.
- Browser
- Download file 7.14 kB
-
https://huggingface.co/datasets/Lilambd/empire-eval/resolve/main/scripts/ds_ground.py
- Command line
-
hf download hf://datasets/Lilambd/empire-eval/scripts/ds_ground.py
-
curl -L -o ds_ground.py https://huggingface.co/datasets/Lilambd/empire-eval/resolve/main/scripts/ds_ground.py
7.14 kB
| DS = 'Lilambd/world-ground' | |
| REVISION = 'hf-main@2026-11-03 (data as of 2026-10-04)' | |
| LICENSE = 'CC BY 4.0' | |
| SH = 'World Ground Ledger' | |
| STR = {'data/admin1_ground.csv': ['a1', 'iso_3166_2', 'adm0_a3'], 'data/features_long.csv': ['a1']} | |
| KEYS = {'data/admin1_ground.csv': ['a1', 'name', 'country'], 'data/features_long.csv': ['a1', 'feature']} | |
| COLS = { | |
| 'pop_volcano50_2020': ('how many people (GHSL 2020) live within 50 km of an erupting volcano', 'persons', {'rel': 0.01}), | |
| 'pop_coast10_2020': ('how many people (GHSL 2020) live within 10 km of the coast', 'persons', {'rel': 0.01}), | |
| 'pop_plate100_2020': ('how many people (GHSL 2020) live within 100 km of a plate boundary', 'persons', {'rel': 0.01}), | |
| 'coast_km_popw': ('what is the population-weighted distance to the coast, in km', 'km', {'rel': 0.01}), | |
| 'plate_km_popw': ('what is the population-weighted distance to a plate boundary, in km', 'km', {'rel': 0.01}), | |
| 'built_2025': ('how many km² are built up in 2025', 'km²', {'rel': 0.01}), | |
| 'soc_0_5': ('what is the soil organic carbon at 0-5 cm, in g/kg', 'g/kg', {'rel': 0.01}), | |
| 'ph_0_5': ('what is the soil pH at 0-5 cm', 'pH', {'abs': 0.01}), | |
| 'ocs_0_30': ('what is the soil organic carbon stock 0-30 cm, in kg/m²', 'kg/m2', {'rel': 0.01}), | |
| } | |
| def build(B): | |
| g = B.store['data/admin1_ground.csv'] | |
| gu = g[~g.duplicated(['name', 'country'], keep=False)] | |
| for col, (txt, u, tol) in COLS.items(): | |
| sub = gu[gu[col].notna()] | |
| if col == 'pop_volcano50_2020': | |
| sub = sub[sub[col] > 0] | |
| for r in sub.loc[B.sample(sub.index, 14)].itertuples(): | |
| B.add('lookup', f'In the {SH}, for {r.name} ({r.country}): {txt}?', {'op': 'lookup', 'file': 'data/admin1_ground.csv', 'where': [['a1', '==', r.a1]], 'col': col}, unit=u, tol=tol, skills=['table_lookup']) | |
| cs = g.groupby('country').size() | |
| for c in B.sample(cs[cs >= 5].index, 40): | |
| col = B.rng.choice(['pop_coast10_2020', 'pop_plate100_2020', 'built_2025', 'coast_km_popw', 'pop_volcano50_2020']) | |
| txt = COLS[col][0].replace('how many people', 'the most people').replace('what is the ', 'the highest ').replace('how many km² are', 'the most km²') | |
| B.add('rank', f'In the {SH}, which admin-1 region of {c} has {txt.replace(", in km", "")}?', | |
| {'op': 'argmax', 'file': 'data/admin1_ground.csv', 'where': [['country', '==', c]], 'col': col, 'return': 'name'}, difficulty=2, skills=['ranking', 'group_filter']) | |
| for c in B.sample(cs[cs >= 3].index, 30): | |
| col = B.rng.choice(['pop_coast10_2020', 'pop_plate100_2020', 'pop_volcano50_2020']) | |
| B.add('aggregate', f'In the {SH}, summing over all admin-1 regions of {c}, {COLS[col][0]}?', | |
| {'op': 'sum', 'file': 'data/admin1_ground.csv', 'where': [['country', '==', c]], 'col': col}, unit='persons', difficulty=2, skills=['aggregation', 'group_filter']) | |
| for c in B.sample(cs[cs >= 5].index, 20): | |
| B.add('count', f'In the {SH}, how many admin-1 regions of {c} have anyone living within 50 km of an erupting volcano (pop_volcano50_2020 > 0)?', | |
| {'op': 'count', 'file': 'data/admin1_ground.csv', 'where': [['country', '==', c], ['pop_volcano50_2020', '>', 0]]}, difficulty=2, skills=['filter', 'count']) | |
| for k in (3, 5): | |
| B.add('rank', f'In the {SH}, list the {k} admin-1 regions with the most people within 50 km of an erupting volcano, most first.', | |
| {'op': 'topk', 'k': k, 'file': 'data/admin1_ground.csv', 'col': 'pop_volcano50_2020', 'return': 'name'}, difficulty=3, skills=['ranking', 'full_scan'], answer_format='JSON list of region names') | |
| B.add('aggregate', f'In the {SH}, how many people in total live within 10 km of the coast (sum of pop_coast10_2020 over all regions)?', {'op': 'sum', 'file': 'data/admin1_ground.csv', 'col': 'pop_coast10_2020'}, unit='persons', difficulty=2, skills=['aggregation', 'full_scan']) | |
| B.add('aggregate', f'In the {SH}, how many people in total live within 50 km of an erupting volcano?', {'op': 'sum', 'file': 'data/admin1_ground.csv', 'col': 'pop_volcano50_2020'}, unit='persons', difficulty=2, skills=['aggregation', 'full_scan']) | |
| for _ in range(30): | |
| a, b = gu.sample(2, random_state=B.rng.randint(0, 1 << 30)).to_dict('records') | |
| col = B.rng.choice(['pop_coast10_2020', 'plate_km_popw', 'built_2025', 'ph_0_5']) | |
| q = {'pop_coast10_2020': 'more people within 10 km of the coast', 'plate_km_popw': 'a larger population-weighted distance to a plate boundary', 'built_2025': 'more built-up km² in 2025', 'ph_0_5': 'a higher topsoil pH'}[col] | |
| B.add('compare', f"In the {SH}, which region has {q}: {a['name']} ({a['country']}) or {b['name']} ({b['country']})?", | |
| {'op': 'compare', 'labels': [a['name'], b['name']], 'parts': [ | |
| {'op': 'lookup', 'file': 'data/admin1_ground.csv', 'where': [['a1', '==', a['a1']]], 'col': col}, | |
| {'op': 'lookup', 'file': 'data/admin1_ground.csv', 'where': [['a1', '==', b['a1']]], 'col': col}]}, difficulty=2, skills=['comparison']) | |
| gi = g[g.iso_3166_2.notna() & ~g.iso_3166_2.astype(str).str.contains('~') & ~g.duplicated(['iso_3166_2'], keep=False)] | |
| for r in gi.loc[B.sample(gi.index, 25)].itertuples(): | |
| col = B.rng.choice(['pop_coast10_2020', 'pop_plate100_2020', 'coast_km_popw']) | |
| B.add('multi_hop', f'In the {SH}, for the region with ISO 3166-2 code {r.iso_3166_2}: {COLS[col][0]}?', | |
| {'op': 'lookup', 'file': 'data/admin1_ground.csv', 'where': [['iso_3166_2', '==', r.iso_3166_2]], 'col': col}, unit=COLS[col][1], difficulty=2, skills=['code_resolution']) | |
| gw = g[g.wikidata.notna() & ~g.duplicated(['wikidata'], keep=False)] | |
| for r in gw.loc[B.sample(gw.index, 15)].itertuples(): | |
| B.add('multi_hop', f'In the {SH}, how many people (GHSL 2020) live within 100 km of a plate boundary in the region with Wikidata id {r.wikidata}?', | |
| {'op': 'lookup', 'file': 'data/admin1_ground.csv', 'where': [['wikidata', '==', r.wikidata]], 'col': 'pop_plate100_2020'}, unit='persons', difficulty=2, skills=['code_resolution']) | |
| soil_na = gu[gu.soc_0_5.isna()] | |
| for r in soil_na.loc[B.sample(soil_na.index, 12)].itertuples(): | |
| B.add('unanswerable', f'In the {SH}, what is the topsoil organic carbon (0-5 cm, g/kg) of {r.name} ({r.country})? If the ledger has no value, answer null.', | |
| {'op': 'absent', 'file': 'data/admin1_ground.csv', 'where': [['a1', '==', r.a1], ['soc_0_5', 'notna', None]]}, difficulty=2, skills=['abstention'], | |
| note='README: about 70 regions have a blank soil value because SoilGrids gave no value for their cells.') | |
| fl = B.store['data/features_long.csv'] | |
| for r in fl.loc[B.sample(fl.index, 20)].itertuples(): | |
| B.add('lookup', f'In the {SH} long table, what is `{r.feature}` ({r.unit}) for region a1={r.a1}?', {'op': 'lookup', 'file': 'data/features_long.csv', 'where': [['a1', '==', r.a1], ['feature', '==', r.feature]], 'col': 'value'}, unit=r.unit, skills=['table_lookup', 'id_lookup']) | |