@@ -374,6 +374,17 @@ def push_ds(ds, path, local=False):
374374 ds .push_to_hub (path )
375375
376376
377+ def get_perf_df (data_dict ):
378+ perfs = {"Model" : []}
379+ for task_id in data_dict [list (data_dict .keys ())[0 ]]:
380+ perfs [task_id ] = []
381+ for model , task_perf in data_dict .items ():
382+ perfs ["Model" ].append (model )
383+ for task_id , status in task_perf .items ():
384+ perfs [task_id ].append (status )
385+ return pd .DataFrame (perfs )
386+
387+
377388if __name__ == "__main__" :
378389
379390 # bcb_orig = load_dataset("bigcode/bigcodebench", split="v0.1.0_hf")
@@ -388,6 +399,9 @@ def push_ds(ds, path, local=False):
388399 files = []
389400 complete_data , complete_files = read_task_perf (bcb ["task_id" ], "complete" )
390401 instruct_data , instruct_files = read_task_perf (bcb ["task_id" ], "instruct" )
402+ complete_df = get_perf_df (complete_data )
403+ instruct_df = get_perf_df (instruct_data )
404+ push_ds (DatasetDict ({"complete" : Dataset .from_pandas (complete_df ), "instruct" : Dataset .from_pandas (instruct_df )}), f"bigcode/bigcodebench{ suffix } -perf" )
391405 assert len (model_info ) == len (complete_data ),\
392406 f"Missing results for { set ([val ['name' ] for val in model_info .values ()]) - set ([model for model in complete_data .keys ()])} "
393407 with open ("task2domain.json" , "r" ) as f :
0 commit comments