File size: 1,543 Bytes
5f3280f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 | """Stage the embed_eval artifacts for an HF-job evaluation (the L4 box is gone).
Reproduces embed_f05.py's singleton sample (duckdb hash order) from parquet/{s1,gt}.parquet and checks that this duckdb
build reproduces the archived eval-query selection of embed_eval.py (same hash function => same singleton sample)."""
import duckdb, pandas as pd, json
E = '/data/amlc/ft/eval'; P = '/data/amlc/parquet'
con = duckdb.connect(); con.execute("SET threads=2; SET memory_limit='6GB'")
ev = pd.read_parquet(f'{E}/embed_eval_set.parquet')
out = {'duckdb': duckdb.__version__}
for c in ['US', 'India']:
q = con.execute(f"""SELECT s1.entity_id FROM read_parquet('{P}/s1.parquet') s1 SEMI JOIN read_parquet('{P}/pairs.parquet') p
ON p.a = s1.entity_id WHERE s1.country = ? ORDER BY hash(s1.entity_id || ?) LIMIT ?""", [c, 'q42', 6000]).df()
arch = set(ev.entity_id[(ev.role == 'query') & (ev.eval_country == c)])
out[f'eval_query_reproduced_{c}'] = f"{len(set(q.entity_id) & arch)}/{len(arch)}"
sing = con.execute(f"""SELECT x.entity_id, x.business_name, x.business_address, x.country
FROM read_parquet('{P}/gt.parquet') g JOIN read_parquet('{P}/s1.parquet') x ON x.entity_id = g.source1_entity_id
WHERE g.matched_entity_ids = '' AND x.country IN ('US','India')
QUALIFY row_number() OVER (PARTITION BY x.country ORDER BY hash(x.entity_id || 'sing7')) <= 400""").df()
sing.to_parquet(f'{E}/f05_singletons.parquet', index=False)
out['singletons'] = sing.country.value_counts().to_dict()
print(json.dumps(out), flush=True)
|