Spaces:
AIR-Bench
/
Running on CPU Upgrade

leaderboard / tests /src /test_benchmarks.py
nan's picture
refactor: refactor the benchmarks
3fcf957
raw
history blame
342 Bytes
from src.benchmarks import BenchmarksQA, BenchmarksLongDoc
def test_qabenchmarks():
print(list(BenchmarksQA))
for benchmark in list(BenchmarksQA):
print(benchmark.name, benchmark.metric, benchmark.col_name, benchmark.domain, benchmark.lang, benchmark.task)
def test_longdocbenchmarks():
print(list(BenchmarksLongDoc))