zhejianzhang commited on
Commit
b72feb2
·
1 Parent(s): 1ea2007
Files changed (1) hide show
  1. app.py +2 -2
app.py CHANGED
@@ -46,10 +46,10 @@ LOCAL_DEBUG=False
46
  TYPES = ["markdown", "number", "number", "number", "number", "str", "str", "str"]
47
 
48
 
49
- benchmark_internal_evaluate_dataset = load_dataset(INTERNAL_DATASET, data_files=BENCHMARK_INTERNAL_EVALUATE_DATASET_FILE, token=TOKEN, verification_mode=VerificationMode.NO_CHECKS, download_mode="force_redownload",trust_remote_code=True)
50
 
51
  print(EVALUATE_RESULT_DATASET_FILE)
52
- eval_results = load_dataset(EVALUATE_RESULT_DATASET, data_files=EVALUATE_RESULT_DATASET_FILE, token=TOKEN, verification_mode=VerificationMode.NO_CHECKS, download_mode="force_redownload",trust_remote_code=True)
53
 
54
  benchmark_dataset = parse_eval_dataset(benchmark_internal_evaluate_dataset) # type: ignore
55
 
 
46
  TYPES = ["markdown", "number", "number", "number", "number", "str", "str", "str"]
47
 
48
 
49
+ benchmark_internal_evaluate_dataset = load_dataset(INTERNAL_DATASET, data_files=BENCHMARK_INTERNAL_EVALUATE_DATASET_FILE, token=TOKEN, verification_mode=VerificationMode.NO_CHECKS, download_mode="reuse_cache_if_exists",trust_remote_code=True)
50
 
51
  print(EVALUATE_RESULT_DATASET_FILE)
52
+ eval_results = load_dataset(EVALUATE_RESULT_DATASET, data_files=EVALUATE_RESULT_DATASET_FILE, token=TOKEN, verification_mode=VerificationMode.NO_CHECKS, download_mode="reuse_cache_if_exists",trust_remote_code=True)
53
 
54
  benchmark_dataset = parse_eval_dataset(benchmark_internal_evaluate_dataset) # type: ignore
55