empire-eval / scripts /common.py
Lilambd's picture
Publish Lilambd Empire Eval
7f9a652 verified
Raw History Blame Contribute Delete
4.44 kB
"""Shared generators reused across Lilambd ledgers."""
import pandas as pd
import pycountry
from evallib import fmt, select
_ISO3 = {}
def cname(iso3):
if iso3 in _ISO3:
return _ISO3[iso3]
c = pycountry.countries.get(alpha_3=iso3)
n = getattr(c, 'common_name', None) or (c.name if c else None)
_ISO3[iso3] = f'{n} ({iso3})' if n else iso3
return _ISO3[iso3]
def cname2(iso2):
c = pycountry.countries.get(alpha_2=iso2)
n = getattr(c, 'common_name', None) or (c.name if c else None)
return f'{n} ({iso2})' if n else iso2
def long_features(B, file, unit_col='unit', key='iso3', label=cname, n_lookup=40, n_rank=12, n_cmp=15, n_count=8, n_absent=10, scope='country', ds_short=''):
"""Questions over a long table (key, feature, value, year, unit)."""
df = B.store[file]
feats = sorted(df['feature'].dropna().unique())
units = df.groupby('feature')[unit_col].first().to_dict()
years = df.groupby('feature')['year'].first().to_dict()
keys_all = sorted(df[key].unique())
per = max(1, n_lookup // max(1, len(feats)))
for ft in feats:
sub = df[df.feature == ft]
u = units[ft]
for _, r in sub.loc[B.sample(sub.index, per)].iterrows():
B.add('lookup', f'In {ds_short} `{file}`, what is the value of `{ft}` ({u}) for {label(r[key])}?',
{'op': 'lookup', 'file': file, 'where': [[key, '==', r[key]], ['feature', '==', ft]], 'col': 'value'},
unit=u, skills=['table_lookup', 'code_resolution'])
for ft in B.sample(feats, n_rank):
u = units[ft]
for op, word in (('argmax', 'highest'), ('argmin', 'lowest')):
B.add('rank', f'In {ds_short} `{file}`, which {scope} has the {word} `{ft}` ({u})? Answer with its {key} code.',
{'op': op, 'file': file, 'where': [['feature', '==', ft]], 'col': 'value', 'return': key},
difficulty=2, skills=['ranking', 'full_scan'])
B.add('rank', f'In {ds_short} `{file}`, list the 3 {scope} codes ({key}) with the highest `{ft}` ({u}), highest first.',
{'op': 'topk', 'k': 3, 'file': file, 'where': [['feature', '==', ft]], 'col': 'value', 'return': key},
difficulty=2, skills=['ranking', 'full_scan'], answer_format='JSON list of codes')
for _ in range(n_cmp):
ft = B.rng.choice(feats)
sub = df[df.feature == ft]
if len(sub) < 2:
continue
a, b = sub.sample(2, random_state=B.rng.randint(0, 1 << 30))[key]
B.add('compare', f'In {ds_short} `{file}`, which has the larger `{ft}` ({units[ft]}): {label(a)} or {label(b)}? Answer with the {key} code.',
{'op': 'compare', 'labels': [a, b], 'parts': [
{'op': 'lookup', 'file': file, 'where': [[key, '==', a], ['feature', '==', ft]], 'col': 'value'},
{'op': 'lookup', 'file': file, 'where': [[key, '==', b], ['feature', '==', ft]], 'col': 'value'}]},
difficulty=2, skills=['comparison'])
for ft in B.sample(feats, n_count):
sub = pd.to_numeric(df[df.feature == ft]['value'], errors='coerce').dropna()
if len(sub) < 10:
continue
thr = float(f'{sub.quantile(B.rng.choice([0.5, 0.75, 0.9])):.3g}')
B.add('count', f'In {ds_short} `{file}`, how many {scope} rows have `{ft}` greater than {fmt(thr)} ({units[ft]})?',
{'op': 'count', 'file': file, 'where': [['feature', '==', ft], ['value', '>', thr]]},
difficulty=2, skills=['filter', 'count', 'full_scan'])
# unanswerable: key exists for other features but not this one
tries = 0
made = 0
while made < n_absent and tries < 200:
tries += 1
ft = B.rng.choice(feats)
have = set(df[df.feature == ft][key])
missing = [k for k in keys_all if k not in have]
if not missing:
continue
k = B.rng.choice(missing)
if B.add('unanswerable', f'In {ds_short} `{file}`, what is the value of `{ft}` ({units[ft]}) for {label(k)}? If the dataset has no such row, answer null.',
{'op': 'absent', 'file': file, 'where': [[key, '==', k], ['feature', '==', ft]]},
difficulty=2, skills=['abstention'],
note='No row exists for this key and feature; a grounded system must abstain.'):
made += 1