evaluation/per_instance.csv

variant,instance_id,success,q_score,steps
human,301,False,0.0,3225
human,302,False,0.0,3225
human,303,True,1.0,1611
human,304,True,1.0,1118
human,305,False,0.0,3225
human,306,False,0.0,3225
human,307,True,1.0,1265
human,308,True,1.0,2112
human,309,False,0.0,3225
human,310,False,0.0,3225
human,311,False,0.0,3225
human,312,True,1.0,1825
human,313,True,1.0,1299
human,314,False,0.0,3225
human,315,True,1.0,1449
human,316,False,0.0,3225
human,317,False,0.0,3225
human,318,False,0.0,3225
human,319,False,0.0,3225
human,320,True,1.0,995
codegen_success,301,False,0.0,3225
codegen_success,302,False,0.0,3225
codegen_success,303,False,0.0,3225
codegen_success,304,False,0.0,3225
codegen_success,305,False,0.0,3225
codegen_success,306,True,1.0,486
codegen_success,307,False,0.0,3225
codegen_success,308,False,0.0,3225
codegen_success,309,False,0.0,3225
codegen_success,310,False,0.0,3225
codegen_success,311,False,0.0,3225
codegen_success,312,False,0.0,3225
codegen_success,313,False,0.0,3225
codegen_success,314,False,0.0,3225
codegen_success,315,False,0.0,3225
codegen_success,316,False,0.0,3225
codegen_success,317,False,0.0,3225
codegen_success,318,False,0.0,3225
codegen_success,319,False,0.0,3225
codegen_success,320,False,0.0,3225
mix_h50_c50,301,False,0.0,3225
mix_h50_c50,302,False,0.0,3225
mix_h50_c50,303,False,0.0,3225
mix_h50_c50,304,False,0.0,3225
mix_h50_c50,305,False,0.0,3225
mix_h50_c50,306,False,0.0,3225
mix_h50_c50,307,False,0.0,3225
mix_h50_c50,308,False,0.0,3225
mix_h50_c50,309,False,0.0,3225
mix_h50_c50,310,False,0.0,3225
mix_h50_c50,311,False,0.0,3225
mix_h50_c50,312,True,1.0,383
mix_h50_c50,313,False,0.0,3225
mix_h50_c50,314,False,0.0,3225
mix_h50_c50,315,False,0.0,3225
mix_h50_c50,316,False,0.0,3225
mix_h50_c50,317,False,0.0,3225
mix_h50_c50,318,False,0.0,3225
mix_h50_c50,319,False,0.0,3225
mix_h50_c50,320,False,0.0,3225