workspace/data/mini_squirrel_results.csv

model,example_id,em,gm,mb,parses,latency_s
qwen2.5-coder:7b,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,13.82
qwen2.5-coder:7b,ecommerce_repeat_buyers__syntax__drop_paren__002,1,1,1,1,3.56
qwen2.5-coder:7b,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,3.15
qwen2.5-coder:7b,ecommerce_repeat_buyers__syntax__bad_column_ref__004,1,1,1,1,3.71
qwen2.5-coder:7b,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,3.74
qwen2.5-coder:7b,ecommerce_repeat_buyers__semantic__swap_aggregate__002,1,1,1,1,3.5
qwen2.5-coder:7b,ecommerce_repeat_buyers__semantic__flip_comparison__003,1,1,1,1,3.56
qwen2.5-coder:7b,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,0,1,3.68
qwen2.5-coder:7b,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,3.53
qwen2.5-coder:7b,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,3.52
qwen2.5-coder:7b,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,3.52
qwen2.5-coder:7b,finance_daily_pnl__syntax__bad_column_ref__008,1,1,1,1,3.93
qwen2.5-coder:7b,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,1,3.67
qwen2.5-coder:7b,finance_daily_pnl__semantic__swap_aggregate__006,0,0,0,1,3.76
qwen2.5-coder:7b,finance_daily_pnl__semantic__flip_comparison__007,1,1,1,1,3.52
qwen2.5-coder:7b,saas_active_seats__syntax__drop_comma__009,1,1,1,1,2.79
qwen2.5-coder:7b,saas_active_seats__syntax__drop_paren__010,1,1,1,1,2.96
qwen2.5-coder:7b,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,3.05
qwen2.5-coder:7b,saas_active_seats__syntax__bad_column_ref__012,1,1,1,1,3.17
qwen2.5-coder:7b,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,3.76
qwen2.5-coder:7b,saas_active_seats__semantic__swap_aggregate__009,0,0,0,1,4.37
qwen2.5-coder:7b,saas_active_seats__semantic__drop_where_predicate__010,0,1,0,1,3.18
qwen2.5-coder:7b,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,3.03
qwen2.5-coder:7b,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,2.87
qwen2.5-coder:7b,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,2.86
qwen2.5-coder:7b,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,3.2
qwen2.5-coder:7b,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,1,3.48
qwen2.5-coder:7b,logistics_late_shipments__semantic__swap_aggregate__012,1,1,1,1,3.17
qwen2.5-coder:7b,logistics_late_shipments__semantic__flip_comparison__013,0,0,0,1,3.55
qwen2.5-coder:7b,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,3.26
qwen2.5-coder:7b,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,3.64
qwen2.5-coder:7b,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,3.53
qwen2.5-coder:7b,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,3.54
qwen2.5-coder:7b,media_engagement_funnel__syntax__bad_column_ref__020,1,1,1,1,3.73
qwen2.5-coder:7b,media_engagement_funnel__semantic__flip_join_type__015,0,0,0,1,3.76
qwen2.5-coder:7b,media_engagement_funnel__semantic__swap_aggregate__016,0,0,0,1,3.66
qwen2.5-coder:7b,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,3.78
qwen2.5-coder:7b,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,0,1,3.45
qwen2.5-coder:7b,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,4.05
qwen2.5-coder:7b,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,4.32
qwen2.5-coder:7b,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,4.66
qwen2.5-coder:7b,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,5.18
qwen2.5-coder:7b,hr_attrition_by_dept__semantic__flip_join_type__019,0,0,0,1,4.96
qwen2.5-coder:7b,hr_attrition_by_dept__semantic__swap_aggregate__020,0,0,0,1,5.21
qwen2.5-coder:7b,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,1,5.04
qwen2.5-coder:7b,hr_attrition_by_dept__semantic__drop_where_predicate__022,0,0,0,1,4.61
qwen2.5-coder:7b,hr_attrition_by_dept__semantic__flip_null_check__023,0,0,0,1,4.97
qwen2.5-coder:7b,iot_sensor_alerts__syntax__drop_comma__025,1,1,1,1,2.71
qwen2.5-coder:7b,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,2.73
qwen2.5-coder:7b,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,2.49
qwen2.5-coder:7b,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,2.49
qwen2.5-coder:7b,iot_sensor_alerts__semantic__flip_join_type__024,1,1,1,1,2.39
qwen2.5-coder:7b,iot_sensor_alerts__semantic__swap_aggregate__025,0,0,0,1,2.48
qwen2.5-coder:7b,iot_sensor_alerts__semantic__flip_comparison__026,1,1,1,1,2.48
qwen2.5-coder:7b,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,3.31
qwen2.5-coder:7b,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,3.34
qwen2.5-coder:7b,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,3.6
qwen2.5-coder:7b,support_ticket_sla__syntax__bad_column_ref__032,1,1,1,1,3.88
qwen2.5-coder:7b,support_ticket_sla__semantic__flip_join_type__027,0,0,0,1,3.89
qwen2.5-coder:7b,support_ticket_sla__semantic__swap_aggregate__028,1,1,1,1,3.91
qwen2.5-coder:7b,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,3.82
qwen2.5-coder:7b,support_ticket_sla__semantic__flip_null_check__030,0,0,0,1,3.84
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,3.89
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__syntax__drop_paren__002,1,1,1,1,1.39
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,1.34
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__syntax__bad_column_ref__004,0,0,1,1,1.51
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,1.57
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__semantic__swap_aggregate__002,0,0,0,1,1.49
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__semantic__flip_comparison__003,1,1,1,1,1.6
qwen2.5-coder:1.5b,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,0,1,1.47
qwen2.5-coder:1.5b,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,1.42
qwen2.5-coder:1.5b,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,1.39
qwen2.5-coder:1.5b,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,1.38
qwen2.5-coder:1.5b,finance_daily_pnl__syntax__bad_column_ref__008,0,0,1,1,1.55
qwen2.5-coder:1.5b,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,1,1.49
qwen2.5-coder:1.5b,finance_daily_pnl__semantic__swap_aggregate__006,0,0,0,1,1.49
qwen2.5-coder:1.5b,finance_daily_pnl__semantic__flip_comparison__007,1,1,1,1,1.66
qwen2.5-coder:1.5b,saas_active_seats__syntax__drop_comma__009,1,1,1,1,1.33
qwen2.5-coder:1.5b,saas_active_seats__syntax__drop_paren__010,0,0,0,0,1.18
qwen2.5-coder:1.5b,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,1.29
qwen2.5-coder:1.5b,saas_active_seats__syntax__bad_column_ref__012,0,0,0,1,1.49
qwen2.5-coder:1.5b,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,1.35
qwen2.5-coder:1.5b,saas_active_seats__semantic__swap_aggregate__009,0,0,0,1,1.29
qwen2.5-coder:1.5b,saas_active_seats__semantic__drop_where_predicate__010,0,0,0,1,1.73
qwen2.5-coder:1.5b,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,1.59
qwen2.5-coder:1.5b,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,1.54
qwen2.5-coder:1.5b,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,1.5
qwen2.5-coder:1.5b,logistics_late_shipments__syntax__bad_column_ref__016,0,0,0,1,1.72
qwen2.5-coder:1.5b,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,1,1.61
qwen2.5-coder:1.5b,logistics_late_shipments__semantic__swap_aggregate__012,0,0,0,1,1.69
qwen2.5-coder:1.5b,logistics_late_shipments__semantic__flip_comparison__013,0,0,0,1,1.61
qwen2.5-coder:1.5b,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,1.56
qwen2.5-coder:1.5b,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,1.72
qwen2.5-coder:1.5b,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,1.74
qwen2.5-coder:1.5b,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,1.9
qwen2.5-coder:1.5b,media_engagement_funnel__syntax__bad_column_ref__020,0,0,0,1,1.86
qwen2.5-coder:1.5b,media_engagement_funnel__semantic__flip_join_type__015,0,0,0,1,1.78
qwen2.5-coder:1.5b,media_engagement_funnel__semantic__swap_aggregate__016,0,0,0,1,1.84
qwen2.5-coder:1.5b,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,1.83
qwen2.5-coder:1.5b,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,0,1,1.68
qwen2.5-coder:1.5b,hr_attrition_by_dept__syntax__drop_comma__021,0,0,0,0,1.86
qwen2.5-coder:1.5b,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,1.95
qwen2.5-coder:1.5b,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,1.87
qwen2.5-coder:1.5b,hr_attrition_by_dept__syntax__bad_column_ref__024,0,0,0,1,2.14
qwen2.5-coder:1.5b,hr_attrition_by_dept__semantic__flip_join_type__019,0,0,0,1,2.0
qwen2.5-coder:1.5b,hr_attrition_by_dept__semantic__swap_aggregate__020,0,0,0,1,2.22
qwen2.5-coder:1.5b,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,1,2.07
qwen2.5-coder:1.5b,hr_attrition_by_dept__semantic__drop_where_predicate__022,0,0,0,1,1.96
qwen2.5-coder:1.5b,hr_attrition_by_dept__semantic__flip_null_check__023,0,0,0,1,2.16
qwen2.5-coder:1.5b,iot_sensor_alerts__syntax__drop_comma__025,0,0,0,0,1.2
qwen2.5-coder:1.5b,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,1.13
qwen2.5-coder:1.5b,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,1.2
qwen2.5-coder:1.5b,iot_sensor_alerts__syntax__bad_column_ref__028,0,0,0,1,1.32
qwen2.5-coder:1.5b,iot_sensor_alerts__semantic__flip_join_type__024,0,0,0,1,1.1
qwen2.5-coder:1.5b,iot_sensor_alerts__semantic__swap_aggregate__025,0,0,0,1,1.17
qwen2.5-coder:1.5b,iot_sensor_alerts__semantic__flip_comparison__026,0,0,0,1,1.15
qwen2.5-coder:1.5b,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,1.73
qwen2.5-coder:1.5b,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,1.63
qwen2.5-coder:1.5b,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,1.7
qwen2.5-coder:1.5b,support_ticket_sla__syntax__bad_column_ref__032,0,0,1,1,1.75
qwen2.5-coder:1.5b,support_ticket_sla__semantic__flip_join_type__027,0,0,0,1,1.82
qwen2.5-coder:1.5b,support_ticket_sla__semantic__swap_aggregate__028,0,0,0,1,1.81
qwen2.5-coder:1.5b,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,1.7
qwen2.5-coder:1.5b,support_ticket_sla__semantic__flip_null_check__030,1,1,1,1,1.87
qwen2.5:7b-instruct,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,9.46
qwen2.5:7b-instruct,ecommerce_repeat_buyers__syntax__drop_paren__002,1,1,1,1,3.7
qwen2.5:7b-instruct,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,3.67
qwen2.5:7b-instruct,ecommerce_repeat_buyers__syntax__bad_column_ref__004,1,1,1,1,4.02
qwen2.5:7b-instruct,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,3.56
qwen2.5:7b-instruct,ecommerce_repeat_buyers__semantic__swap_aggregate__002,1,1,1,1,4.01
qwen2.5:7b-instruct,ecommerce_repeat_buyers__semantic__flip_comparison__003,1,1,1,1,4.03
qwen2.5:7b-instruct,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,0,1,3.8
qwen2.5:7b-instruct,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,3.65
qwen2.5:7b-instruct,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,3.59
qwen2.5:7b-instruct,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,3.5
qwen2.5:7b-instruct,finance_daily_pnl__syntax__bad_column_ref__008,1,1,1,1,3.44
qwen2.5:7b-instruct,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,1,3.46
qwen2.5:7b-instruct,finance_daily_pnl__semantic__swap_aggregate__006,0,0,0,1,3.42
qwen2.5:7b-instruct,finance_daily_pnl__semantic__flip_comparison__007,1,1,1,1,3.38
qwen2.5:7b-instruct,saas_active_seats__syntax__drop_comma__009,1,1,1,1,2.66
qwen2.5:7b-instruct,saas_active_seats__syntax__drop_paren__010,1,1,1,1,2.74
qwen2.5:7b-instruct,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,2.7
qwen2.5:7b-instruct,saas_active_seats__syntax__bad_column_ref__012,1,1,1,1,2.84
qwen2.5:7b-instruct,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,3.55
qwen2.5:7b-instruct,saas_active_seats__semantic__swap_aggregate__009,1,1,1,1,2.85
qwen2.5:7b-instruct,saas_active_seats__semantic__drop_where_predicate__010,1,1,1,1,2.85
qwen2.5:7b-instruct,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,2.96
qwen2.5:7b-instruct,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,2.93
qwen2.5:7b-instruct,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,2.98
qwen2.5:7b-instruct,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,3.2
qwen2.5:7b-instruct,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,1,3.21
qwen2.5:7b-instruct,logistics_late_shipments__semantic__swap_aggregate__012,1,1,1,1,3.12
qwen2.5:7b-instruct,logistics_late_shipments__semantic__flip_comparison__013,1,1,1,1,3.3
qwen2.5:7b-instruct,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,3.17
qwen2.5:7b-instruct,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,3.62
qwen2.5:7b-instruct,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,3.71
qwen2.5:7b-instruct,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,3.87
qwen2.5:7b-instruct,media_engagement_funnel__syntax__bad_column_ref__020,1,1,1,1,4.43
qwen2.5:7b-instruct,media_engagement_funnel__semantic__flip_join_type__015,0,0,0,1,4.29
qwen2.5:7b-instruct,media_engagement_funnel__semantic__swap_aggregate__016,0,0,0,1,3.86
qwen2.5:7b-instruct,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,3.86
qwen2.5:7b-instruct,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,0,1,3.76
qwen2.5:7b-instruct,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,4.11
qwen2.5:7b-instruct,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,3.99
qwen2.5:7b-instruct,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,4.01
qwen2.5:7b-instruct,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,4.66
qwen2.5:7b-instruct,hr_attrition_by_dept__semantic__flip_join_type__019,0,0,0,1,4.23
qwen2.5:7b-instruct,hr_attrition_by_dept__semantic__swap_aggregate__020,0,0,0,1,4.69
qwen2.5:7b-instruct,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,1,4.47
qwen2.5:7b-instruct,hr_attrition_by_dept__semantic__drop_where_predicate__022,0,0,0,1,4.31
qwen2.5:7b-instruct,hr_attrition_by_dept__semantic__flip_null_check__023,1,1,1,1,4.68
qwen2.5:7b-instruct,iot_sensor_alerts__syntax__drop_comma__025,1,1,1,1,2.41
qwen2.5:7b-instruct,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,2.44
qwen2.5:7b-instruct,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,2.4
qwen2.5:7b-instruct,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,2.83
qwen2.5:7b-instruct,iot_sensor_alerts__semantic__flip_join_type__024,0,0,0,1,2.52
qwen2.5:7b-instruct,iot_sensor_alerts__semantic__swap_aggregate__025,0,0,0,1,2.59
qwen2.5:7b-instruct,iot_sensor_alerts__semantic__flip_comparison__026,0,1,1,1,2.82
qwen2.5:7b-instruct,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,3.69
qwen2.5:7b-instruct,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,3.67
qwen2.5:7b-instruct,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,3.38
qwen2.5:7b-instruct,support_ticket_sla__syntax__bad_column_ref__032,1,1,1,1,3.53
qwen2.5:7b-instruct,support_ticket_sla__semantic__flip_join_type__027,0,0,0,1,3.42
qwen2.5:7b-instruct,support_ticket_sla__semantic__swap_aggregate__028,0,0,0,1,4.04
qwen2.5:7b-instruct,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,3.86
qwen2.5:7b-instruct,support_ticket_sla__semantic__flip_null_check__030,1,1,1,1,3.69
__identity__,ecommerce_repeat_buyers__syntax__drop_comma__001,0,0,0,0,0.0
__identity__,ecommerce_repeat_buyers__syntax__drop_paren__002,0,0,0,0,0.0
__identity__,ecommerce_repeat_buyers__syntax__misspell_keyword__003,0,0,0,0,0.0
__identity__,ecommerce_repeat_buyers__syntax__bad_column_ref__004,0,0,0,1,0.0
__identity__,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,0.0
__identity__,ecommerce_repeat_buyers__semantic__swap_aggregate__002,0,0,0,1,0.0
__identity__,ecommerce_repeat_buyers__semantic__flip_comparison__003,0,0,0,1,0.0
__identity__,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,0,1,0.0
__identity__,finance_daily_pnl__syntax__drop_comma__005,0,0,0,0,0.0
__identity__,finance_daily_pnl__syntax__drop_paren__006,0,0,0,0,0.0
__identity__,finance_daily_pnl__syntax__misspell_keyword__007,0,0,0,0,0.0
__identity__,finance_daily_pnl__syntax__bad_column_ref__008,0,0,0,1,0.0
__identity__,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,1,0.0
__identity__,finance_daily_pnl__semantic__swap_aggregate__006,0,0,0,1,0.0
__identity__,finance_daily_pnl__semantic__flip_comparison__007,0,0,0,1,0.0
__identity__,saas_active_seats__syntax__drop_comma__009,0,0,0,0,0.0
__identity__,saas_active_seats__syntax__drop_paren__010,0,0,0,0,0.0
__identity__,saas_active_seats__syntax__misspell_keyword__011,0,0,0,0,0.0
__identity__,saas_active_seats__syntax__bad_column_ref__012,0,0,0,1,0.0
__identity__,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,0.0
__identity__,saas_active_seats__semantic__swap_aggregate__009,0,0,0,1,0.0
__identity__,saas_active_seats__semantic__drop_where_predicate__010,0,0,0,1,0.0
__identity__,logistics_late_shipments__syntax__drop_comma__013,0,0,0,0,0.0
__identity__,logistics_late_shipments__syntax__drop_paren__014,0,0,0,0,0.0
__identity__,logistics_late_shipments__syntax__misspell_keyword__015,0,0,0,0,0.0
__identity__,logistics_late_shipments__syntax__bad_column_ref__016,0,0,0,1,0.0
__identity__,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,1,0.0
__identity__,logistics_late_shipments__semantic__swap_aggregate__012,0,0,0,1,0.0
__identity__,logistics_late_shipments__semantic__flip_comparison__013,0,0,0,1,0.0
__identity__,logistics_late_shipments__semantic__flip_null_check__014,0,0,0,1,0.0
__identity__,media_engagement_funnel__syntax__drop_comma__017,0,0,0,0,0.0
__identity__,media_engagement_funnel__syntax__drop_paren__018,0,0,0,0,0.0
__identity__,media_engagement_funnel__syntax__misspell_keyword__019,0,0,0,0,0.0
__identity__,media_engagement_funnel__syntax__bad_column_ref__020,0,0,0,1,0.0
__identity__,media_engagement_funnel__semantic__flip_join_type__015,0,0,0,1,0.0
__identity__,media_engagement_funnel__semantic__swap_aggregate__016,0,0,0,1,0.0
__identity__,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,0.0
__identity__,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,0,1,0.0
__identity__,hr_attrition_by_dept__syntax__drop_comma__021,0,0,0,0,0.0
__identity__,hr_attrition_by_dept__syntax__drop_paren__022,0,0,0,0,0.0
__identity__,hr_attrition_by_dept__syntax__misspell_keyword__023,0,0,0,0,0.0
__identity__,hr_attrition_by_dept__syntax__bad_column_ref__024,0,0,0,1,0.0
__identity__,hr_attrition_by_dept__semantic__flip_join_type__019,0,0,0,1,0.0
__identity__,hr_attrition_by_dept__semantic__swap_aggregate__020,0,0,0,1,0.0
__identity__,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,1,0.0
__identity__,hr_attrition_by_dept__semantic__drop_where_predicate__022,0,0,0,1,0.0
__identity__,hr_attrition_by_dept__semantic__flip_null_check__023,0,0,0,1,0.0
__identity__,iot_sensor_alerts__syntax__drop_comma__025,0,0,0,0,0.0
__identity__,iot_sensor_alerts__syntax__drop_paren__026,0,0,0,0,0.0
__identity__,iot_sensor_alerts__syntax__misspell_keyword__027,0,0,0,0,0.0
__identity__,iot_sensor_alerts__syntax__bad_column_ref__028,0,0,0,1,0.0
__identity__,iot_sensor_alerts__semantic__flip_join_type__024,0,0,0,1,0.0
__identity__,iot_sensor_alerts__semantic__swap_aggregate__025,0,0,0,1,0.0
__identity__,iot_sensor_alerts__semantic__flip_comparison__026,0,0,0,1,0.0
__identity__,support_ticket_sla__syntax__drop_comma__029,0,0,0,0,0.0
__identity__,support_ticket_sla__syntax__drop_paren__030,0,0,0,0,0.0
__identity__,support_ticket_sla__syntax__misspell_keyword__031,0,0,0,0,0.0
__identity__,support_ticket_sla__syntax__bad_column_ref__032,0,0,0,1,0.0
__identity__,support_ticket_sla__semantic__flip_join_type__027,0,0,0,1,0.0
__identity__,support_ticket_sla__semantic__swap_aggregate__028,0,0,0,1,0.0
__identity__,support_ticket_sla__semantic__flip_comparison__029,0,0,0,1,0.0
__identity__,support_ticket_sla__semantic__flip_null_check__030,0,0,0,1,0.0
claude-sonnet-4.5,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,3.02
claude-sonnet-4.5,ecommerce_repeat_buyers__syntax__drop_paren__002,1,1,1,1,3.2
claude-sonnet-4.5,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,3.57
claude-sonnet-4.5,ecommerce_repeat_buyers__syntax__bad_column_ref__004,1,1,1,1,3.2
claude-sonnet-4.5,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,3.16
claude-sonnet-4.5,ecommerce_repeat_buyers__semantic__swap_aggregate__002,1,1,1,1,2.96
claude-sonnet-4.5,ecommerce_repeat_buyers__semantic__flip_comparison__003,1,1,1,1,3.19
claude-sonnet-4.5,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,1,1,3.36
claude-sonnet-4.5,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,3.45
claude-sonnet-4.5,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,2.93
claude-sonnet-4.5,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,3.18
claude-sonnet-4.5,finance_daily_pnl__syntax__bad_column_ref__008,1,1,1,1,7.45
claude-sonnet-4.5,finance_daily_pnl__semantic__flip_join_type__005,0,1,0,1,3.48
claude-sonnet-4.5,finance_daily_pnl__semantic__swap_aggregate__006,1,1,1,1,2.86
claude-sonnet-4.5,finance_daily_pnl__semantic__flip_comparison__007,1,1,1,1,3.05
claude-sonnet-4.5,saas_active_seats__syntax__drop_comma__009,1,1,1,1,2.3
claude-sonnet-4.5,saas_active_seats__syntax__drop_paren__010,1,1,1,1,2.33
claude-sonnet-4.5,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,2.17
claude-sonnet-4.5,saas_active_seats__syntax__bad_column_ref__012,1,1,1,1,2.47
claude-sonnet-4.5,saas_active_seats__semantic__flip_join_type__008,0,1,0,1,2.78
claude-sonnet-4.5,saas_active_seats__semantic__swap_aggregate__009,1,1,1,1,2.89
claude-sonnet-4.5,saas_active_seats__semantic__drop_where_predicate__010,1,1,1,1,3.22
claude-sonnet-4.5,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,3.19
claude-sonnet-4.5,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,2.47
claude-sonnet-4.5,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,2.59
claude-sonnet-4.5,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,3.14
claude-sonnet-4.5,logistics_late_shipments__semantic__flip_join_type__011,1,1,1,1,5.81
claude-sonnet-4.5,logistics_late_shipments__semantic__swap_aggregate__012,1,1,1,1,2.99
claude-sonnet-4.5,logistics_late_shipments__semantic__flip_comparison__013,1,1,1,1,2.48
claude-sonnet-4.5,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,2.49
claude-sonnet-4.5,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,2.69
claude-sonnet-4.5,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,3.15
claude-sonnet-4.5,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,2.75
claude-sonnet-4.5,media_engagement_funnel__syntax__bad_column_ref__020,1,1,1,1,3.39
claude-sonnet-4.5,media_engagement_funnel__semantic__flip_join_type__015,0,1,0,1,3.56
claude-sonnet-4.5,media_engagement_funnel__semantic__swap_aggregate__016,1,1,1,1,2.92
claude-sonnet-4.5,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,6.18
claude-sonnet-4.5,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,1,1,2.97
claude-sonnet-4.5,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,3.36
claude-sonnet-4.5,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,2.91
claude-sonnet-4.5,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,4.03
claude-sonnet-4.5,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,3.63
claude-sonnet-4.5,hr_attrition_by_dept__semantic__flip_join_type__019,1,1,1,1,3.59
claude-sonnet-4.5,hr_attrition_by_dept__semantic__swap_aggregate__020,1,1,1,1,3.02
claude-sonnet-4.5,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,1,4.09
claude-sonnet-4.5,hr_attrition_by_dept__semantic__drop_where_predicate__022,1,1,1,1,3.14
claude-sonnet-4.5,hr_attrition_by_dept__semantic__flip_null_check__023,0,0,0,1,6.19
claude-sonnet-4.5,iot_sensor_alerts__syntax__drop_comma__025,1,1,1,1,2.24
claude-sonnet-4.5,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,2.61
claude-sonnet-4.5,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,1.92
claude-sonnet-4.5,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,2.69
claude-sonnet-4.5,iot_sensor_alerts__semantic__flip_join_type__024,1,1,1,1,9.94
claude-sonnet-4.5,iot_sensor_alerts__semantic__swap_aggregate__025,1,1,1,1,2.53
claude-sonnet-4.5,iot_sensor_alerts__semantic__flip_comparison__026,1,1,1,1,2.82
claude-sonnet-4.5,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,2.82
claude-sonnet-4.5,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,2.51
claude-sonnet-4.5,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,2.62
claude-sonnet-4.5,support_ticket_sla__syntax__bad_column_ref__032,1,1,1,1,3.15
claude-sonnet-4.5,support_ticket_sla__semantic__flip_join_type__027,0,1,0,1,2.69
claude-sonnet-4.5,support_ticket_sla__semantic__swap_aggregate__028,1,1,1,1,3.6
claude-sonnet-4.5,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,3.51
claude-sonnet-4.5,support_ticket_sla__semantic__flip_null_check__030,1,1,1,1,2.71
claude-sonnet-4.6,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,3.26
claude-sonnet-4.6,ecommerce_repeat_buyers__syntax__drop_paren__002,1,1,1,1,3.15
claude-sonnet-4.6,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,4.81
claude-sonnet-4.6,ecommerce_repeat_buyers__syntax__bad_column_ref__004,1,1,1,1,3.11
claude-sonnet-4.6,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,9.2
claude-sonnet-4.6,ecommerce_repeat_buyers__semantic__swap_aggregate__002,1,1,1,1,3.82
claude-sonnet-4.6,ecommerce_repeat_buyers__semantic__flip_comparison__003,1,1,1,1,3.24
claude-sonnet-4.6,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,1,1,6.83
claude-sonnet-4.6,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,3.14
claude-sonnet-4.6,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,3.16
claude-sonnet-4.6,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,3.01
claude-sonnet-4.6,finance_daily_pnl__syntax__bad_column_ref__008,1,1,1,1,12.47
claude-sonnet-4.6,finance_daily_pnl__semantic__flip_join_type__005,0,1,0,1,9.02
claude-sonnet-4.6,finance_daily_pnl__semantic__swap_aggregate__006,1,1,1,1,4.84
claude-sonnet-4.6,finance_daily_pnl__semantic__flip_comparison__007,1,1,1,1,3.74
claude-sonnet-4.6,saas_active_seats__syntax__drop_comma__009,1,1,1,1,3.04
claude-sonnet-4.6,saas_active_seats__syntax__drop_paren__010,1,1,1,1,2.46
claude-sonnet-4.6,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,2.58
claude-sonnet-4.6,saas_active_seats__syntax__bad_column_ref__012,1,1,1,1,2.75
claude-sonnet-4.6,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,8.19
claude-sonnet-4.6,saas_active_seats__semantic__swap_aggregate__009,1,1,1,1,3.37
claude-sonnet-4.6,saas_active_seats__semantic__drop_where_predicate__010,1,1,1,1,2.8
claude-sonnet-4.6,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,2.74
claude-sonnet-4.6,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,2.69
claude-sonnet-4.6,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,10.13
claude-sonnet-4.6,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,2.65
claude-sonnet-4.6,logistics_late_shipments__semantic__flip_join_type__011,0,1,0,1,8.06
claude-sonnet-4.6,logistics_late_shipments__semantic__swap_aggregate__012,1,1,1,1,7.41
claude-sonnet-4.6,logistics_late_shipments__semantic__flip_comparison__013,1,1,1,1,4.33
claude-sonnet-4.6,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,4.02
claude-sonnet-4.6,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,5.13
claude-sonnet-4.6,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,3.29
claude-sonnet-4.6,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,2.89
claude-sonnet-4.6,media_engagement_funnel__syntax__bad_column_ref__020,1,1,1,1,3.64
claude-sonnet-4.6,media_engagement_funnel__semantic__flip_join_type__015,0,1,0,1,8.98
claude-sonnet-4.6,media_engagement_funnel__semantic__swap_aggregate__016,1,1,1,1,9.9
claude-sonnet-4.6,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,5.29
claude-sonnet-4.6,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,0,0,6.15
claude-sonnet-4.6,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,3.02
claude-sonnet-4.6,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,10.41
claude-sonnet-4.6,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,4.09
claude-sonnet-4.6,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,3.33
claude-sonnet-4.6,hr_attrition_by_dept__semantic__flip_join_type__019,1,1,1,1,4.64
claude-sonnet-4.6,hr_attrition_by_dept__semantic__swap_aggregate__020,1,1,1,1,4.67
claude-sonnet-4.6,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,0,5.39
claude-sonnet-4.6,hr_attrition_by_dept__semantic__drop_where_predicate__022,1,1,1,1,5.75
claude-sonnet-4.6,hr_attrition_by_dept__semantic__flip_null_check__023,0,0,0,0,5.62
claude-sonnet-4.6,iot_sensor_alerts__syntax__drop_comma__025,1,1,1,1,2.24
claude-sonnet-4.6,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,4.69
claude-sonnet-4.6,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,2.25
claude-sonnet-4.6,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,2.58
claude-sonnet-4.6,iot_sensor_alerts__semantic__flip_join_type__024,1,1,1,1,13.61
claude-sonnet-4.6,iot_sensor_alerts__semantic__swap_aggregate__025,1,1,1,1,2.26
claude-sonnet-4.6,iot_sensor_alerts__semantic__flip_comparison__026,1,1,1,1,3.18
claude-sonnet-4.6,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,3.14
claude-sonnet-4.6,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,3.03
claude-sonnet-4.6,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,4.55
claude-sonnet-4.6,support_ticket_sla__syntax__bad_column_ref__032,1,1,1,1,2.88
claude-sonnet-4.6,support_ticket_sla__semantic__flip_join_type__027,0,1,0,1,4.86
claude-sonnet-4.6,support_ticket_sla__semantic__swap_aggregate__028,0,1,0,1,3.07
claude-sonnet-4.6,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,2.99
claude-sonnet-4.6,support_ticket_sla__semantic__flip_null_check__030,1,1,1,1,4.57
claude-sonnet-5,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,3.34
claude-sonnet-5,ecommerce_repeat_buyers__syntax__drop_paren__002,1,1,1,1,2.66
claude-sonnet-5,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,3.07
claude-sonnet-5,ecommerce_repeat_buyers__syntax__bad_column_ref__004,1,1,1,1,2.91
claude-sonnet-5,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,0,12.4
claude-sonnet-5,ecommerce_repeat_buyers__semantic__swap_aggregate__002,1,1,1,1,2.8
claude-sonnet-5,ecommerce_repeat_buyers__semantic__flip_comparison__003,1,1,1,1,2.99
claude-sonnet-5,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,1,1,1,1,3.34
claude-sonnet-5,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,3.57
claude-sonnet-5,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,2.81
claude-sonnet-5,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,2.67
claude-sonnet-5,finance_daily_pnl__syntax__bad_column_ref__008,1,1,1,1,8.44
claude-sonnet-5,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,0,11.96
claude-sonnet-5,finance_daily_pnl__semantic__swap_aggregate__006,1,1,1,1,5.99
claude-sonnet-5,finance_daily_pnl__semantic__flip_comparison__007,1,1,1,1,2.81
claude-sonnet-5,saas_active_seats__syntax__drop_comma__009,1,1,1,1,2.49
claude-sonnet-5,saas_active_seats__syntax__drop_paren__010,1,1,1,1,2.33
claude-sonnet-5,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,2.49
claude-sonnet-5,saas_active_seats__syntax__bad_column_ref__012,1,1,1,1,2.4
claude-sonnet-5,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,6.99
claude-sonnet-5,saas_active_seats__semantic__swap_aggregate__009,1,1,1,1,2.49
claude-sonnet-5,saas_active_seats__semantic__drop_where_predicate__010,1,1,1,1,2.41
claude-sonnet-5,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,2.97
claude-sonnet-5,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,2.88
claude-sonnet-5,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,4.2
claude-sonnet-5,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,2.53
claude-sonnet-5,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,0,11.67
claude-sonnet-5,logistics_late_shipments__semantic__swap_aggregate__012,1,1,1,1,3.3
claude-sonnet-5,logistics_late_shipments__semantic__flip_comparison__013,1,1,1,1,2.64
claude-sonnet-5,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,2.56
claude-sonnet-5,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,3.33
claude-sonnet-5,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,2.88
claude-sonnet-5,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,3.05
claude-sonnet-5,media_engagement_funnel__syntax__bad_column_ref__020,1,1,1,1,3.01
claude-sonnet-5,media_engagement_funnel__semantic__flip_join_type__015,0,1,0,1,3.98
claude-sonnet-5,media_engagement_funnel__semantic__swap_aggregate__016,0,0,0,0,11.8
claude-sonnet-5,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,4.08
claude-sonnet-5,media_engagement_funnel__semantic__drop_where_predicate__018,1,1,1,1,4.94
claude-sonnet-5,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,3.24
claude-sonnet-5,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,3.64
claude-sonnet-5,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,3.19
claude-sonnet-5,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,3.03
claude-sonnet-5,hr_attrition_by_dept__semantic__flip_join_type__019,1,1,1,1,3.46
claude-sonnet-5,hr_attrition_by_dept__semantic__swap_aggregate__020,1,1,1,1,4.36
claude-sonnet-5,hr_attrition_by_dept__semantic__flip_comparison__021,1,1,1,1,3.41
claude-sonnet-5,hr_attrition_by_dept__semantic__drop_where_predicate__022,1,1,1,1,4.09
claude-sonnet-5,hr_attrition_by_dept__semantic__flip_null_check__023,1,1,1,1,3.58
claude-sonnet-5,iot_sensor_alerts__syntax__drop_comma__025,1,1,1,1,3.36
claude-sonnet-5,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,3.03
claude-sonnet-5,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,2.6
claude-sonnet-5,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,2.25
claude-sonnet-5,iot_sensor_alerts__semantic__flip_join_type__024,0,0,0,1,6.05
claude-sonnet-5,iot_sensor_alerts__semantic__swap_aggregate__025,1,1,1,1,2.22
claude-sonnet-5,iot_sensor_alerts__semantic__flip_comparison__026,1,1,1,1,2.18
claude-sonnet-5,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,3.51
claude-sonnet-5,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,6.87
claude-sonnet-5,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,2.76
claude-sonnet-5,support_ticket_sla__syntax__bad_column_ref__032,1,1,1,1,2.95
claude-sonnet-5,support_ticket_sla__semantic__flip_join_type__027,0,1,0,1,5.25
claude-sonnet-5,support_ticket_sla__semantic__swap_aggregate__028,1,1,1,1,3.01
claude-sonnet-5,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,2.91
claude-sonnet-5,support_ticket_sla__semantic__flip_null_check__030,1,1,1,1,3.76
gpt-5,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,8.58
gpt-5,ecommerce_repeat_buyers__syntax__drop_paren__002,1,1,1,1,5.68
gpt-5,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,6.06
gpt-5,ecommerce_repeat_buyers__syntax__bad_column_ref__004,1,1,1,1,5.61
gpt-5,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,38.99
gpt-5,ecommerce_repeat_buyers__semantic__swap_aggregate__002,0,0,0,1,15.75
gpt-5,ecommerce_repeat_buyers__semantic__flip_comparison__003,0,0,0,1,13.35
gpt-5,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,0,1,10.95
gpt-5,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,3.57
gpt-5,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,10.64
gpt-5,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,5.28
gpt-5,finance_daily_pnl__syntax__bad_column_ref__008,1,1,1,1,4.71
gpt-5,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,1,27.32
gpt-5,finance_daily_pnl__semantic__swap_aggregate__006,0,0,0,1,15.9
gpt-5,finance_daily_pnl__semantic__flip_comparison__007,1,1,1,1,8.2
gpt-5,saas_active_seats__syntax__drop_comma__009,1,1,1,1,5.36
gpt-5,saas_active_seats__syntax__drop_paren__010,1,1,1,1,4.42
gpt-5,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,4.22
gpt-5,saas_active_seats__syntax__bad_column_ref__012,1,1,1,1,3.13
gpt-5,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,14.86
gpt-5,saas_active_seats__semantic__swap_aggregate__009,0,0,0,1,15.98
gpt-5,saas_active_seats__semantic__drop_where_predicate__010,0,0,0,1,20.21
gpt-5,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,5.29
gpt-5,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,5.57
gpt-5,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,5.09
gpt-5,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,4.55
gpt-5,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,1,21.6
gpt-5,logistics_late_shipments__semantic__swap_aggregate__012,0,1,0,1,12.07
gpt-5,logistics_late_shipments__semantic__flip_comparison__013,0,1,0,1,12.14
gpt-5,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,10.82
gpt-5,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,11.12
gpt-5,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,6.29
gpt-5,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,9.55
gpt-5,media_engagement_funnel__syntax__bad_column_ref__020,1,1,1,1,10.47
gpt-5,media_engagement_funnel__semantic__flip_join_type__015,0,0,0,1,37.73
gpt-5,media_engagement_funnel__semantic__swap_aggregate__016,1,1,1,1,22.46
gpt-5,media_engagement_funnel__semantic__flip_comparison__017,1,1,1,1,19.68
gpt-5,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,1,1,26.47
gpt-5,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,8.37
gpt-5,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,8.41
gpt-5,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,8.4
gpt-5,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,8.39
gpt-5,hr_attrition_by_dept__semantic__flip_join_type__019,1,1,1,1,26.4
gpt-5,hr_attrition_by_dept__semantic__swap_aggregate__020,0,0,0,1,30.69
gpt-5,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,1,40.46
gpt-5,hr_attrition_by_dept__semantic__drop_where_predicate__022,1,1,1,1,25.72
gpt-5,hr_attrition_by_dept__semantic__flip_null_check__023,0,0,0,1,27.82
gpt-5,iot_sensor_alerts__syntax__drop_comma__025,1,1,1,1,3.71
gpt-5,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,6.19
gpt-5,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,3.05
gpt-5,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,4.7
gpt-5,iot_sensor_alerts__semantic__flip_join_type__024,0,0,0,1,26.36
gpt-5,iot_sensor_alerts__semantic__swap_aggregate__025,0,1,0,1,9.45
gpt-5,iot_sensor_alerts__semantic__flip_comparison__026,1,1,1,1,6.65
gpt-5,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,7.14
gpt-5,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,6.86
gpt-5,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,9.13
gpt-5,support_ticket_sla__syntax__bad_column_ref__032,1,1,1,1,7.04
gpt-5,support_ticket_sla__semantic__flip_join_type__027,0,0,0,1,26.28
gpt-5,support_ticket_sla__semantic__swap_aggregate__028,1,1,1,1,9.61
gpt-5,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,10.53
gpt-5,support_ticket_sla__semantic__flip_null_check__030,0,1,0,1,10.44
gemini-3.1-pro-preview,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,41.44
gemini-3.1-pro-preview,ecommerce_repeat_buyers__syntax__drop_paren__002,0,0,0,0,0.0
gemini-3.1-pro-preview,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,46.03
gemini-3.1-pro-preview,ecommerce_repeat_buyers__syntax__bad_column_ref__004,1,1,1,1,117.62
gemini-3.1-pro-preview,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,1,53.07
gemini-3.1-pro-preview,ecommerce_repeat_buyers__semantic__swap_aggregate__002,0,0,0,1,36.41
gemini-3.1-pro-preview,ecommerce_repeat_buyers__semantic__flip_comparison__003,1,1,1,1,37.45
gemini-3.1-pro-preview,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,1,1,48.05
gemini-3.1-pro-preview,finance_daily_pnl__syntax__drop_comma__005,1,1,1,1,64.93
gemini-3.1-pro-preview,finance_daily_pnl__syntax__drop_paren__006,1,1,1,1,33.99
gemini-3.1-pro-preview,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,48.11
gemini-3.1-pro-preview,finance_daily_pnl__syntax__bad_column_ref__008,1,1,1,1,38.85
gemini-3.1-pro-preview,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,0,60.21
gemini-3.1-pro-preview,finance_daily_pnl__semantic__swap_aggregate__006,1,1,1,1,33.32
gemini-3.1-pro-preview,finance_daily_pnl__semantic__flip_comparison__007,0,0,0,0,0.0
gemini-3.1-pro-preview,saas_active_seats__syntax__drop_comma__009,1,1,1,1,130.01
gemini-3.1-pro-preview,saas_active_seats__syntax__drop_paren__010,1,1,1,1,34.87
gemini-3.1-pro-preview,saas_active_seats__syntax__misspell_keyword__011,1,1,1,1,52.14
gemini-3.1-pro-preview,saas_active_seats__syntax__bad_column_ref__012,1,1,1,1,40.44
gemini-3.1-pro-preview,saas_active_seats__semantic__flip_join_type__008,0,0,0,1,69.39
gemini-3.1-pro-preview,saas_active_seats__semantic__swap_aggregate__009,1,1,1,1,79.75
gemini-3.1-pro-preview,saas_active_seats__semantic__drop_where_predicate__010,1,1,1,1,41.44
gemini-3.1-pro-preview,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,51.78
gemini-3.1-pro-preview,logistics_late_shipments__syntax__drop_paren__014,1,1,1,1,43.72
gemini-3.1-pro-preview,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,45.82
gemini-3.1-pro-preview,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,40.31
gemini-3.1-pro-preview,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,0,72.54
gemini-3.1-pro-preview,logistics_late_shipments__semantic__swap_aggregate__012,1,1,1,1,31.38
gemini-3.1-pro-preview,logistics_late_shipments__semantic__flip_comparison__013,1,1,1,1,31.66
gemini-3.1-pro-preview,logistics_late_shipments__semantic__flip_null_check__014,1,1,1,1,29.35
gemini-3.1-pro-preview,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,22.16
gemini-3.1-pro-preview,media_engagement_funnel__syntax__drop_paren__018,1,1,1,1,21.52
gemini-3.1-pro-preview,media_engagement_funnel__syntax__misspell_keyword__019,1,1,1,1,38.32
gemini-3.1-pro-preview,media_engagement_funnel__syntax__bad_column_ref__020,1,1,1,1,22.21
gemini-3.1-pro-preview,media_engagement_funnel__semantic__flip_join_type__015,0,0,0,0,45.67
gemini-3.1-pro-preview,media_engagement_funnel__semantic__swap_aggregate__016,1,1,1,1,52.69
gemini-3.1-pro-preview,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,23.43
gemini-3.1-pro-preview,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,1,1,34.12
gemini-3.1-pro-preview,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,51.78
gemini-3.1-pro-preview,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,18.05
gemini-3.1-pro-preview,hr_attrition_by_dept__syntax__misspell_keyword__023,1,1,1,1,19.36
gemini-3.1-pro-preview,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,23.82
gemini-3.1-pro-preview,hr_attrition_by_dept__semantic__flip_join_type__019,1,1,1,1,23.84
gemini-3.1-pro-preview,hr_attrition_by_dept__semantic__swap_aggregate__020,0,0,0,0,97.12
gemini-3.1-pro-preview,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,1,28.68
gemini-3.1-pro-preview,hr_attrition_by_dept__semantic__drop_where_predicate__022,1,1,1,1,27.84
gemini-3.1-pro-preview,hr_attrition_by_dept__semantic__flip_null_check__023,1,1,1,1,24.13
gemini-3.1-pro-preview,iot_sensor_alerts__syntax__drop_comma__025,1,1,1,1,40.71
gemini-3.1-pro-preview,iot_sensor_alerts__syntax__drop_paren__026,1,1,1,1,63.66
gemini-3.1-pro-preview,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,5.99
gemini-3.1-pro-preview,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,4.34
gemini-3.1-pro-preview,iot_sensor_alerts__semantic__flip_join_type__024,0,0,0,1,10.83
gemini-3.1-pro-preview,iot_sensor_alerts__semantic__swap_aggregate__025,1,1,1,1,7.19
gemini-3.1-pro-preview,iot_sensor_alerts__semantic__flip_comparison__026,1,1,1,1,5.23
gemini-3.1-pro-preview,support_ticket_sla__syntax__drop_comma__029,1,1,1,1,7.15
gemini-3.1-pro-preview,support_ticket_sla__syntax__drop_paren__030,1,1,1,1,8.36
gemini-3.1-pro-preview,support_ticket_sla__syntax__misspell_keyword__031,1,1,1,1,8.95
gemini-3.1-pro-preview,support_ticket_sla__syntax__bad_column_ref__032,1,1,1,1,5.07
gemini-3.1-pro-preview,support_ticket_sla__semantic__flip_join_type__027,0,0,0,0,32.32
gemini-3.1-pro-preview,support_ticket_sla__semantic__swap_aggregate__028,1,1,1,1,6.72
gemini-3.1-pro-preview,support_ticket_sla__semantic__flip_comparison__029,1,1,1,1,7.54
gemini-3.1-pro-preview,support_ticket_sla__semantic__flip_null_check__030,1,1,1,1,7.4
deepseek-r1:8b,ecommerce_repeat_buyers__syntax__drop_comma__001,1,1,1,1,241.24
deepseek-r1:8b,ecommerce_repeat_buyers__syntax__drop_paren__002,0,0,0,0,0.0
deepseek-r1:8b,ecommerce_repeat_buyers__syntax__misspell_keyword__003,1,1,1,1,29.61
deepseek-r1:8b,ecommerce_repeat_buyers__syntax__bad_column_ref__004,0,0,1,1,115.61
deepseek-r1:8b,ecommerce_repeat_buyers__semantic__flip_join_type__001,0,0,0,0,0.0
deepseek-r1:8b,ecommerce_repeat_buyers__semantic__swap_aggregate__002,0,0,0,1,99.82
deepseek-r1:8b,ecommerce_repeat_buyers__semantic__flip_comparison__003,0,0,1,1,161.8
deepseek-r1:8b,ecommerce_repeat_buyers__semantic__drop_where_predicate__004,0,0,0,1,222.23
deepseek-r1:8b,finance_daily_pnl__syntax__drop_comma__005,0,0,0,0,0.0
deepseek-r1:8b,finance_daily_pnl__syntax__drop_paren__006,0,1,1,1,185.94
deepseek-r1:8b,finance_daily_pnl__syntax__misspell_keyword__007,1,1,1,1,120.22
deepseek-r1:8b,finance_daily_pnl__syntax__bad_column_ref__008,0,0,0,0,0.0
deepseek-r1:8b,finance_daily_pnl__semantic__flip_join_type__005,0,0,0,0,0.0
deepseek-r1:8b,finance_daily_pnl__semantic__swap_aggregate__006,1,1,1,1,141.5
deepseek-r1:8b,finance_daily_pnl__semantic__flip_comparison__007,0,0,1,1,150.47
deepseek-r1:8b,saas_active_seats__syntax__drop_comma__009,0,0,0,0,0.0
deepseek-r1:8b,saas_active_seats__syntax__drop_paren__010,1,1,1,1,32.12
deepseek-r1:8b,saas_active_seats__syntax__misspell_keyword__011,0,0,1,1,54.79
deepseek-r1:8b,saas_active_seats__syntax__bad_column_ref__012,0,0,0,1,37.24
deepseek-r1:8b,saas_active_seats__semantic__flip_join_type__008,0,0,0,0,0.0
deepseek-r1:8b,saas_active_seats__semantic__swap_aggregate__009,0,0,0,0,0.0
deepseek-r1:8b,saas_active_seats__semantic__drop_where_predicate__010,0,0,0,0,0.0
deepseek-r1:8b,logistics_late_shipments__syntax__drop_comma__013,1,1,1,1,130.4
deepseek-r1:8b,logistics_late_shipments__syntax__drop_paren__014,0,0,0,0,0.0
deepseek-r1:8b,logistics_late_shipments__syntax__misspell_keyword__015,1,1,1,1,27.99
deepseek-r1:8b,logistics_late_shipments__syntax__bad_column_ref__016,1,1,1,1,16.68
deepseek-r1:8b,logistics_late_shipments__semantic__flip_join_type__011,0,0,0,0,0.0
deepseek-r1:8b,logistics_late_shipments__semantic__swap_aggregate__012,0,0,0,0,0.0
deepseek-r1:8b,logistics_late_shipments__semantic__flip_comparison__013,0,0,0,1,128.24
deepseek-r1:8b,logistics_late_shipments__semantic__flip_null_check__014,0,0,0,1,226.93
deepseek-r1:8b,media_engagement_funnel__syntax__drop_comma__017,1,1,1,1,126.01
deepseek-r1:8b,media_engagement_funnel__syntax__drop_paren__018,0,1,1,1,238.05
deepseek-r1:8b,media_engagement_funnel__syntax__misspell_keyword__019,0,0,0,0,0.0
deepseek-r1:8b,media_engagement_funnel__syntax__bad_column_ref__020,0,0,0,1,80.99
deepseek-r1:8b,media_engagement_funnel__semantic__flip_join_type__015,0,0,0,1,217.67
deepseek-r1:8b,media_engagement_funnel__semantic__swap_aggregate__016,0,0,0,0,0.0
deepseek-r1:8b,media_engagement_funnel__semantic__flip_comparison__017,0,0,0,1,240.48
deepseek-r1:8b,media_engagement_funnel__semantic__drop_where_predicate__018,0,0,0,0,0.0
deepseek-r1:8b,hr_attrition_by_dept__syntax__drop_comma__021,1,1,1,1,31.67
deepseek-r1:8b,hr_attrition_by_dept__syntax__drop_paren__022,1,1,1,1,182.68
deepseek-r1:8b,hr_attrition_by_dept__syntax__misspell_keyword__023,0,0,0,0,0.0
deepseek-r1:8b,hr_attrition_by_dept__syntax__bad_column_ref__024,1,1,1,1,101.79
deepseek-r1:8b,hr_attrition_by_dept__semantic__flip_join_type__019,0,0,0,0,0.0
deepseek-r1:8b,hr_attrition_by_dept__semantic__swap_aggregate__020,0,0,0,0,0.0
deepseek-r1:8b,hr_attrition_by_dept__semantic__flip_comparison__021,0,0,0,0,0.0
deepseek-r1:8b,hr_attrition_by_dept__semantic__drop_where_predicate__022,0,0,0,0,0.0
deepseek-r1:8b,hr_attrition_by_dept__semantic__flip_null_check__023,0,0,0,0,0.0
deepseek-r1:8b,iot_sensor_alerts__syntax__drop_comma__025,0,0,0,0,0.0
deepseek-r1:8b,iot_sensor_alerts__syntax__drop_paren__026,0,0,0,0,0.0
deepseek-r1:8b,iot_sensor_alerts__syntax__misspell_keyword__027,1,1,1,1,220.7
deepseek-r1:8b,iot_sensor_alerts__syntax__bad_column_ref__028,1,1,1,1,41.34
deepseek-r1:8b,iot_sensor_alerts__semantic__flip_join_type__024,0,0,0,0,0.0
deepseek-r1:8b,iot_sensor_alerts__semantic__swap_aggregate__025,0,0,0,0,0.0
deepseek-r1:8b,iot_sensor_alerts__semantic__flip_comparison__026,0,1,1,1,102.56
deepseek-r1:8b,support_ticket_sla__syntax__drop_comma__029,0,0,0,0,0.0
deepseek-r1:8b,support_ticket_sla__syntax__drop_paren__030,0,0,0,0,0.0
deepseek-r1:8b,support_ticket_sla__syntax__misspell_keyword__031,0,0,0,0,0.0
deepseek-r1:8b,support_ticket_sla__syntax__bad_column_ref__032,0,0,0,0,140.95
deepseek-r1:8b,support_ticket_sla__semantic__flip_join_type__027,0,0,0,0,0.0
deepseek-r1:8b,support_ticket_sla__semantic__swap_aggregate__028,0,0,0,0,0.0
deepseek-r1:8b,support_ticket_sla__semantic__flip_comparison__029,0,0,0,1,214.71
deepseek-r1:8b,support_ticket_sla__semantic__flip_null_check__030,0,0,0,1,124.56