empire-eval / scripts /ds_pulse.py
Lilambd's picture
Publish Lilambd Empire Eval
7f9a652 verified
Raw History Blame Contribute Delete
5.41 kB
from remote_util import latest, day_of, uniq
DS = 'Lilambd/world-pulse-daily'
LICENSE = 'CC BY-SA 4.0'
STR = {}
KEYS = {}
SH = 'World Pulse Daily'
# table -> list of (where-col, value-col, phrase) lookups keyed by `key`
SPEC = {
'fx_ecb': [('rate', 'what is the ECB reference rate of EUR to {name}')],
'arxiv': [('submitted', 'how many arXiv submissions are counted for category {name}')],
'stablecoins': [('circulating', 'what circulating stablecoin amount (USD) is recorded for {name}')],
'hackernews_front': [('score', 'what score does the Hacker News front-page story ranked {name} have'), ('title', 'what is the title of the Hacker News front-page story ranked {name}')],
'wikipedia_size': [('articles', 'how many articles does the {name} Wikipedia have'), ('activeusers', 'how many active users does the {name} Wikipedia have')],
'docker_official_images': [('pulls', 'how many pulls does the Docker official image {name} have'), ('stars', 'how many stars does the Docker official image {name} have')],
'stackoverflow_tags': [('questions', 'how many questions are counted for the Stack Overflow key {name}')],
'ai_sdk_downloads': [('downloads', 'how many downloads are recorded for the AI SDK package key {name}')],
'crossref': [('dois', 'how many DOIs are counted for the Crossref facet key {name}')],
'zenodo': [('records', 'how many records are counted for the Zenodo facet key {name}')],
'hf_daily_papers': [('upvotes', 'how many upvotes does the Hugging Face daily paper with key {name} have')],
'github_new_repos': [('repos', 'how many new GitHub repos are counted for the key {name}')],
'crates_io': [('downloads', 'how many downloads does the crates.io key {name} show')],
'wikipedia_top': [('article', 'which article holds the Wikipedia top-views key {name}'), ('views', 'how many views does the Wikipedia top-views key {name} have')],
'appstore_charts': [('name', 'which app holds the App Store chart position key {name}')],
}
def build(B):
root = B.store.root
for t, specs in SPEC.items():
f = latest(root, f'{t}/*.csv')
if not f:
continue
KEYS[f] = ['key']
d = B.store[f]
day = day_of(f)
du = uniq(d, 'key')
n = 30 if t in ('appstore_charts', 'wikipedia_top', 'stablecoins') else 14
for r in du.loc[B.sample(du.index, n)].to_dict('records'):
col, phrase = B.rng.choice(specs)
if col not in r or r[col] != r[col] or r[col] is None:
continue
nm = r['key'] if t not in ('fx_ecb',) else r['key']
B.add('lookup', f'In {SH} ({t}, fetched {day}), {phrase.format(name=nm)}?', {'op': 'lookup', 'file': f, 'where': [['key', '==', r['key']]], 'col': col},
tol={'rel': 0.0001}, skills=['table_lookup', 'fresh_data'])
num = [c for c in ('submitted', 'circulating', 'score', 'articles', 'pulls', 'questions', 'downloads', 'dois', 'records', 'upvotes', 'repos', 'views') if c in d.columns]
for c in num[:1]:
B.add('rank', f'In {SH} ({t}, fetched {day}), which key has the highest {c}?', {'op': 'argmax', 'file': f, 'col': c, 'return': 'key'}, difficulty=2, skills=['ranking', 'fresh_data'])
B.add('rank', f'In {SH} ({t}, fetched {day}), list the 3 keys with the highest {c}, highest first.', {'op': 'topk', 'k': 3, 'file': f, 'col': c, 'return': 'key'}, difficulty=2, skills=['ranking', 'fresh_data'], answer_format='JSON list')
B.add('count', f'In {SH}, how many rows does the {t} table have for fetch date {day}?', {'op': 'count', 'file': f}, skills=['count', 'fresh_data'])
f = latest(root, 'appstore_charts/*.csv')
if f:
a = B.store[f]
for c in B.sample(sorted(a.country.dropna().unique()), 25):
ch = B.rng.choice(sorted(a.chart.dropna().unique()))
B.add('lookup', f'In {SH} App Store charts fetched {day_of(f)}, which app is #1 on the {ch} chart in country "{c}"?', {'op': 'lookup', 'file': f, 'where': [['country', '==', c], ['chart', '==', ch], ['rank', '==', 1]], 'col': 'name'}, skills=['table_lookup', 'fresh_data', 'multilingual_entity'])
for dev in list(a.developer.dropna().value_counts().index)[:12]:
B.add('count', f'In {SH} App Store charts fetched {day_of(f)}, how many chart entries (all countries, all charts) belong to the developer "{dev}"?', {'op': 'count', 'file': f, 'where': [['developer', '==', dev]]}, difficulty=2, skills=['filter', 'count', 'fresh_data'])
f = latest(root, 'wikipedia_top/*.csv')
if f:
w = B.store[f]
for lg in B.sample(sorted(w.lang.dropna().unique()), 15):
rk = B.rng.choice([3, 4, 5, 10])
B.add('lookup', f'In {SH} Wikipedia top views fetched {day_of(f)}, which article is ranked #{rk} in the "{lg}" Wikipedia?', {'op': 'lookup', 'file': f, 'where': [['lang', '==', lg], ['rank', '==', rk]], 'col': 'article'}, skills=['table_lookup', 'fresh_data', 'multilingual_entity'])
f = latest(root, 'fx_ecb/*.csv')
if f:
for cur in ['XYZ', 'LIL']:
B.add('unanswerable', f'In {SH} ECB rates fetched {day_of(f)}, what is the EUR rate for the currency code {cur}? If not listed, answer null.', {'op': 'absent', 'file': f, 'where': [['currency', '==', cur]]}, skills=['abstention'])