Datasets:
File size: 1,848 Bytes
baddcf2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 | """Deterministic source/task-balanced training pilot; no evaluation examples."""
import argparse,json
from pathlib import Path
import duckdb,pyarrow as pa
from assemble import allocate
def main():
p=argparse.ArgumentParser();p.add_argument('--root',required=True);p.add_argument('--output',required=True);p.add_argument('--rows',type=int,default=1000);a=p.parse_args()
db=duckdb.connect();db.execute("SET memory_limit='16GB'")
pattern=str(Path(a.root)/'selected'/'part-*.parquet')
db.execute("CREATE TABLE pool AS SELECT *,provenance.repository AS source FROM read_parquet(?) WHERE split='train' AND array_length(options) BETWEEN 2 AND 20",[pattern])
available=dict(db.execute('SELECT source,count(*) FROM pool GROUP BY source').fetchall())
quotas=allocate(available,a.rows,{k:1 for k in available});caps=[]
for source,quota in quotas.items():
strata=dict(db.execute('SELECT task,count(*) FROM pool WHERE source=? GROUP BY task',[source]).fetchall())
for task,cap in allocate(strata,quota,{k:1 for k in strata}).items():caps.append({'source':source,'task':task,'cap':cap})
db.register('caps',pa.Table.from_pylist(caps))
db.execute('CREATE TABLE picked AS SELECT p.* FROM pool p JOIN caps c USING(source,task) QUALIFY row_number() OVER(PARTITION BY source,task ORDER BY sha256(id))<=c.cap')
# Round-robin source order means the first 19 records touch all 19 sources.
table=db.execute('SELECT * EXCLUDE(source) FROM picked QUALIFY true ORDER BY row_number() OVER(PARTITION BY source ORDER BY sha256(id)),source').fetch_arrow_table()
with open(a.output,'w') as f:
for row in table.to_pylist():f.write(json.dumps(row,ensure_ascii=False)+'\n')
print(json.dumps({'rows':len(table),'sources':quotas,'split':'train','max_options':20}),flush=True)
if __name__=='__main__':main()
|