empire-eval / scripts /ds_signals.py
Lilambd's picture
Publish Lilambd Empire Eval
7f9a652 verified
Raw History Blame Contribute Delete
8.81 kB
from remote_util import latest, day_of, has, uniq
from common import cname2
DS = 'Lilambd/world-signals'
LICENSE = 'CC BY 4.0'
STR = {}
KEYS = {}
SH = 'World Signals'
def clean(B, f):
"""True if the file exists and has no exact duplicate rows. The 2026-11-03
export repeated rows (gdelt_events: 12,013 rows, 871 distinct); counts and
rankings over such a file would be wrong, so a file with duplicates is skipped."""
if not latest(B.store.root, f):
return False
d = B.store[f]
dup = int(d.drop(columns=['_row']).duplicated().sum())
if dup:
print(f' skip {f}: {dup} exact duplicate rows')
return False
return True
def build(B):
root = B.store.root
day = day_of(latest(root, 'data/*/steam_top.csv'))
if not day:
return
on = f'In {SH} for {day}'
f = f'data/{day}/steam_top.csv'
if clean(B, f):
KEYS[f] = ['rank', 'appid', 'name']
st = B.store[f]
for r in uniq(st, 'name').head(60).pipe(lambda d: d.loc[B.sample(d.index, 25)]).itertuples():
col = B.rng.choice(['rank', 'peak_players'])
B.add('lookup', f'{on}, ' + (f'what rank does "{r.name}" hold in the Steam top list?' if col == 'rank' else f'what is the peak concurrent player count of "{r.name}"?'),
{'op': 'lookup', 'file': f, 'where': [['name', '==', r.name]], 'col': col}, tol={'abs': 0}, skills=['table_lookup', 'fresh_data'])
for k in (1, 3, 5):
B.add('rank', f'{on}, which {k} Steam games have the highest peak players, highest first?' if k > 1 else f'{on}, which Steam game has the highest peak players?',
{'op': 'topk' if k > 1 else 'argmax', 'k': k, 'file': f, 'col': 'peak_players', 'return': 'name'}, difficulty=2, skills=['ranking', 'fresh_data'], answer_format='JSON list' if k > 1 else None)
for g in ['Free To Play', 'Indie', 'RPG', 'Strategy', 'Simulation', 'Action']:
B.add('count', f'{on}, how many games in the Steam top list have the genre "{g}"?', {'op': 'count', 'file': f, 'where': [['genres', 'contains', g]]}, difficulty=2, skills=['filter', 'count', 'fresh_data'])
f = f'data/{day}/github_new_repos.csv'
if clean(B, f):
KEYS[f] = ['repo']
g = B.store[f]
for r in uniq(g, 'repo').pipe(lambda d: d.loc[B.sample(d.index, 25)]).itertuples():
col = B.rng.choice(['stars', 'forks', 'language'])
if col == 'language' and r.language != r.language:
continue
B.add('lookup', f'{on}, how many {col} does the new GitHub repo {r.repo} have?' if col != 'language' else f'{on}, what is the main language of the new GitHub repo {r.repo}?',
{'op': 'lookup', 'file': f, 'where': [['repo', '==', r.repo]], 'col': col}, tol={'abs': 0}, skills=['table_lookup', 'fresh_data'])
langs = g.language.value_counts()
for lg in list(langs[langs >= 3].index)[:12]:
B.add('count', f'{on}, how many of the new GitHub repos are written mainly in {lg}?', {'op': 'count', 'file': f, 'where': [['language', '==', lg]]}, skills=['filter', 'count', 'fresh_data'])
B.add('rank', f'{on}, which new GitHub repo in {lg} has the most stars?', {'op': 'argmax', 'file': f, 'where': [['language', '==', lg]], 'col': 'stars', 'return': 'repo'}, difficulty=2, skills=['ranking', 'fresh_data'])
f = f'data/{day}/hf_trending.csv'
if clean(B, f):
KEYS[f] = ['kind', 'repo_id']
h = B.store[f]
for r in uniq(h, 'repo_id').pipe(lambda d: d.loc[B.sample(d.index, 25)]).itertuples():
col = B.rng.choice(['likes', 'downloads_month', 'trending_score'])
B.add('lookup', f'{on}, what is the {col.replace("_", " ")} of the Hugging Face repo {r.repo_id}?', {'op': 'lookup', 'file': f, 'where': [['repo_id', '==', r.repo_id]], 'col': col}, tol={'abs': 0}, skills=['table_lookup', 'fresh_data'])
for k in sorted(h.kind.dropna().unique()):
B.add('rank', f'{on}, which {k.replace("hf_", "Hugging Face ")} has the highest trending score?', {'op': 'argmax', 'file': f, 'where': [['kind', '==', k]], 'col': 'trending_score', 'return': 'repo_id'}, difficulty=2, skills=['ranking', 'fresh_data'])
B.add('count', f'{on}, how many trending {k.replace("hf_", "Hugging Face ")}s are listed?', {'op': 'count', 'file': f, 'where': [['kind', '==', k]]}, skills=['count', 'fresh_data'])
for p in list(h.pipeline.dropna().value_counts().index)[:8]:
B.add('count', f'{on}, how many trending Hugging Face models have the pipeline "{p}"?', {'op': 'count', 'file': f, 'where': [['pipeline', '==', p]]}, skills=['filter', 'count', 'fresh_data'])
f = f'data/{day}/package_downloads.csv'
if clean(B, f):
KEYS[f] = ['ecosystem', 'package']
pk = B.store[f]
pku = pk[~pk.duplicated(['ecosystem', 'package'], keep=False) & pk.downloads_month.notna()]
for r in pku.loc[B.sample(pku.index, 30)].itertuples():
B.add('lookup', f'{on}, how many monthly downloads does the {r.ecosystem} package "{r.package}" have?', {'op': 'lookup', 'file': f, 'where': [['ecosystem', '==', r.ecosystem], ['package', '==', r.package]], 'col': 'downloads_month'}, tol={'rel': 0.0001}, skills=['table_lookup', 'fresh_data'])
for e in sorted(pk.ecosystem.dropna().unique()):
B.add('rank', f'{on}, which {e} package has the most monthly downloads?', {'op': 'argmax', 'file': f, 'where': [['ecosystem', '==', e]], 'col': 'downloads_month', 'return': 'package'}, difficulty=2, skills=['ranking', 'fresh_data'])
B.add('rank', f'{on}, list the 3 {e} packages with the most monthly downloads, most first.', {'op': 'topk', 'k': 3, 'file': f, 'where': [['ecosystem', '==', e]], 'col': 'downloads_month', 'return': 'package'}, difficulty=2, skills=['ranking', 'fresh_data'], answer_format='JSON list')
for _ in range(10):
a, b = pku.sample(2, random_state=B.rng.randint(0, 1 << 30)).to_dict('records')
B.add('compare', f"{on}, which package has more monthly downloads: {a['package']} ({a['ecosystem']}) or {b['package']} ({b['ecosystem']})?",
{'op': 'compare', 'labels': [a['package'], b['package']], 'parts': [
{'op': 'lookup', 'file': f, 'where': [['ecosystem', '==', a['ecosystem']], ['package', '==', a['package']]], 'col': 'downloads_month'},
{'op': 'lookup', 'file': f, 'where': [['ecosystem', '==', b['ecosystem']], ['package', '==', b['package']]], 'col': 'downloads_month'}]}, difficulty=2, skills=['comparison', 'fresh_data'])
f = f'data/{day}/podcast_charts.csv'
if clean(B, f):
KEYS[f] = ['country', 'rank', 'name']
pc = B.store[f]
for c in B.sample(sorted(pc.country.dropna().unique()), 25):
rk = B.rng.choice([1, 1, 2, 3, 5, 10])
B.add('lookup', f'{on}, which podcast is ranked #{rk} on the Apple Podcasts chart for {cname2(c.upper())}?', {'op': 'lookup', 'file': f, 'where': [['country', '==', c], ['rank', '==', rk]], 'col': 'name'}, skills=['table_lookup', 'fresh_data', 'multilingual_entity'])
for c in B.sample(sorted(pc.country.dropna().unique()), 10):
B.add('rank', f'{on}, which genre appears most often on the Apple Podcasts chart for {cname2(c.upper())}?', {'op': 'group_top', 'file': f, 'where': [['country', '==', c]], 'by': 'genre'}, difficulty=3, skills=['group_by', 'fresh_data'])
f = f'data/{day}/search_trends.csv'
if clean(B, f):
KEYS[f] = ['country', 'term']
s = B.store[f]
su = s[~s.duplicated(['country', 'term'], keep=False)]
for r in su.loc[B.sample(su.index, 20)].itertuples():
B.add('lookup', f'{on}, what approximate search traffic is shown for the trending term "{r.term}" in {cname2(str(r.country).upper())}?', {'op': 'lookup', 'file': f, 'where': [['country', '==', r.country], ['term', '==', r.term]], 'col': 'approx_traffic'}, tol={'abs': 0}, skills=['table_lookup', 'fresh_data', 'multilingual_entity'])
for c in B.sample(sorted(s.country.dropna().unique()), 12):
B.add('count', f'{on}, how many trending search terms are listed for {cname2(c.upper())}?', {'op': 'count', 'file': f, 'where': [['country', '==', c]]}, skills=['count', 'fresh_data'])
for nm in ['Lilambd Quest Online', 'Hibe Riverfolk Simulator']:
f = f'data/{day}/steam_top.csv'
if clean(B, f):
B.add('unanswerable', f'{on}, what rank does "{nm}" hold in the Steam top list? If it is not listed, answer null.', {'op': 'absent', 'file': f, 'where': [['name', '==', nm]]}, skills=['abstention', 'fresh_data'])