Spaces:
AIR-Bench
/
Running on CPU Upgrade

leaderboard / tests /src /test_benchmarks.py
nan's picture
feat: implement the version selector for qa
7845083
raw
history blame
432 Bytes
from src.benchmarks import QABenchmarks, LongDocBenchmarks
def test_qabenchmarks():
for benchmark_list in list(QABenchmarks):
print(benchmark_list.name)
for b in list(benchmark_list.value):
print(b)
qa_benchmarks = QABenchmarks["2404"]
l = list(frozenset([c.value.domain for c in list(qa_benchmarks.value)]))
print(l)
def test_longdocbenchmarks():
print(list(LongDocBenchmarks))