Datasets:
task_id stringlengths 9 16 | id stringlengths 31 39 | name stringlengths 7 145 | workload stringclasses 2
values | engine stringclasses 5
values | category stringclasses 5
values | language stringclasses 2
values | modality stringclasses 1
value | timeout_seconds int64 600 1.8k | prompt stringlengths 386 5.82k | ground_truth stringlengths 317 30k | expected_csv stringlengths 0 5.39k | grade_spec_csv stringlengths 0 1.01k | path stringlengths 37 48 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
hivesql_001_en | offline-compute_HiveSQL_hivesql_001 | Message Queue Topic Dimension Table internal_platform_db.dim_mq_topic_d_su | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
**Task Objective**: Read data from the previous day's partition of the input table and copy it as-is to the output table's current day partition.
**Time Variables**: The platform provides these variables for dynamic date computation:
- `${yyyymmdd}` : current day in YYYYMMDD format
- `${yyyymmdd-1}` : previo... | INSERT overwrite TABLE internal_platform_db.dim_mq_topic_d_copilot_query_engine_001 PARTITION (dt = '20260507')
SELECT
business_id
,business_name
,cluster_set
,tenant
,namespaces
,topic
,mq_type
,dw_appgroup
,in_charge
,description
,create_time
,modify_time
,cluster_id
,cluster_type
,clust... | business_id,business_name,cluster_set,tenant,namespaces,topic,mq_type,dw_appgroup,in_charge,description,create_time,modify_time,cluster_id,cluster_type,cluster_name,bg,category_name,is_filtered,tids,consumed_tids,unconsumed_tids,is_fully_consumed,has_unconsumed_tid,system_belong,dt
bid001,BizName1,cluster_set_a,tenant_... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,25列(5) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,25,非key字段逐列值匹配率
D,D_field_completeness,field_completeness,15,关键字段非空/非空串比例
F,F_insert_overwrite,insert_overwrite,... | tasks/offline-compute/HiveSQL/hivesql_001_en |
hivesql_002 | offline-compute_HiveSQL_hivesql_002 | 统计数据平台WDNotebook Ray类型管道任务中运行时间跨自然天的实例明细。从`wedat | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
**任务目标**:统计数据平台WDNotebook Ray类型管道任务中运行时间跨自然天的实例明细,按自然天拆分并关联GPU指标。
**输入**:
- `internal_platform_db.notebook_span_info_query_engine_005`
- `internal_platform_db.dwd_gputj_service_instance_map_query_engine_005`
- `internal_platform_db.dwd_ml_platform_instance_podname_query_engine_005`
- `internal_platform_db.gp... | INSERT overwrite TABLE internal_platform_db.dwd_notebook_instance_pod_cross_day_detail_d_query_engine_005 PARTITION (dt = '20260507')
WITH
base_data_raw AS (
SELECT DISTINCT
trace_id,
span_name,
start_time,
end_time,
datawd_project_id,
datawd_task_id,
datawd_t... | trace_id,datawd_project_id,datawd_task_id,datawd_task_instance_id,compute_type,status_code,instance_run_time,code_run_time,resource_wait_time,code_start_time,code_end_time,instance_start_time,instance_end_time,serving_id,is_permanent,apply_for_gpu_count,pod_name,pkg_agg_time,gpu_util,gpu_count,p_date,dt
trace_001,proj_... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,22列 + 列名匹配
C,C_row_alignment,row_consistency,10,行数比例 + key覆盖率
D,D_time_calculation,time_calculation,30,"instance_run_time, code_run_time 等时间指标"
D,D_cross_day_split,cross_day_split,15,trace_001 应出现在两天中
D,D_dimension_join,dimens... | tasks/offline-compute/HiveSQL/hivesql_002 |
hivesql_003 | offline-compute_HiveSQL_hivesql_003 | 从 `internal_platform_db.notebook_span_info_query_engine | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
1) 任务目标
识别并统计 Notebook 管道中被异常 kill 且缺失正常完成信号的 Ray 实例,按天拆分计算运行时长、代码执行时长、资源等待时长及 GPU 申请数。
2) 输入
- `internal_platform_db.notebook_span_info_query_engine_007`
- `internal_platform_db.notebook_engine_info_query_engine_007`
3) 处理规则
- 目标实例筛选:从 span_info 筛选 `compute_type='ray'` 且 `service_name='notebook-runner'` 的实... | INSERT overwrite TABLE internal_platform_db.dwd_notebook_killed_instance_detail_d_copilot_query_engine_007 PARTITION (dt = '20260507')
WITH base_trace AS (
SELECT trace_id
FROM internal_platform_db.notebook_span_info_query_engine_007
WHERE databus_imp_date >= '2026050700'
AND databus_imp_date <= '2026... | p_date,trace_id,datawd_project_id,datawd_task_id,datawd_task_instance_id,compute_type,status_code,instance_run_time,code_run_time,resource_wait_time,code_start_time,code_end_time,instance_start_time,instance_end_time,serving_id,is_permanent,apply_for_gpu_count,dt
2025-05-07,trace_001,proj_01,task_01,inst_01,ray,2,15,13... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,18列 + 列名匹配
C,C_row_alignment,row_consistency,10,行数比例 + key覆盖率
D,D_time_calculation,time_calculation,30,"instance_run_time, code_run_time 等时间指标"
D,D_cross_day_split,cross_day_split,15,跨天 trace 按天拆分正确性
D,D_engine_join,engine_joi... | tasks/offline-compute/HiveSQL/hivesql_003 |
hivesql_004 | offline-compute_HiveSQL_hivesql_004 | 将源表 internal_platform_db.app_group_product_info_sup | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:将应用组与产品归属关系按天做历史快照,按分区全量覆盖写入目标表。
输入:internal_platform_db.app_group_product_info_query_engine_011
- 字段:username、application_group、application_group_owner、product_id、product_name、product_owner、obs_product_id、obs_product_name、obs_product_owner、department、plan_product_id、plan_product_name、plan_product_owner... | insert overwrite TABLE internal_platform_db.app_group_product_info_history_query_engine_011 PARTITION (dt = '20260507')
SELECT
username,
application_group,
application_group_owner,
product_id,
product_name,
product_owner,
obs_product_id,
obs_product_name,
obs_product_owner,
department,
plan_product_id,
plan_product_nam... | username,application_group,application_group_owner,product_id,product_name,product_owner,obs_product_id,obs_product_name,obs_product_owner,department,plan_product_id,plan_product_name,plan_product_owner,application_group_numbers,application_group_owners,product_owners,obs_product_owners,plan_product_owners,bg,id,dt
use... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,21列(5) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,25,非key字段逐列值匹配率
D,D_field_completeness,field_completeness,15,关键字段非空/非空串比例
F,F_insert_overwrite,insert_overwrite,... | tasks/offline-compute/HiveSQL/hivesql_004 |
hivesql_005 | offline-compute_HiveSQL_hivesql_005 | 从 internal_platform_db.dws_mq_production_featur | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:汇总当天有生产量的 消息队列MQ topic 维度信息及近7/30/90天生产统计,落地为 topic 治理项明细表。
输入:
- 表:internal_platform_db.dws_mq_production_feature_d_increase_query_engine_013
- 分区字段:dt(STRING,格式 YYYYMMDD)
- 关键字段:business_id、business_name、topic、cluster_set、tenant、namespaces、system_belong、dw_appgroup、in_charge、description、create_time、mo... | INSERT OVERWRITE TABLE internal_platform_db.ads_mq_topic_governance_item_d_query_engine_013 PARTITION (dt = '20260507')
SELECT
business_id,
business_name,
topic,
cluster_set,
IF ( tenant IS NOT NULL AND namespaces IS NOT NULL, CONCAT( 'persistent://', tenant, '/', namespaces, '/', topic ), topic ) AS mq_full_... | business_id,business_name,topic,cluster_set,mq_full_topic,system_belong,app_group,bid_incharge,bid_description,bid_create_time,bid_modify_time,cluster_id,cluster_type,cluster_name,bg,category_name,total_produce_pkg_d,production_days_last_7d,total_produce_pkg_last_7d,production_days_last_30d,total_produce_pkg_last_30d,p... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,30列(5) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_mapping,field_mapping,25,字段别名映射正确性(app_group/bid_*等)
D,D_derived_mq_full_topic,derived_mq_full_topic,15,mq_full_topic 派生列拼接逻辑正确性
F,F_insert_ov... | tasks/offline-compute/HiveSQL/hivesql_005 |
hivesql_006_en | offline-compute_HiveSQL_hivesql_006 | Input Table internal_platform_db.ods_t_databus_access_topic | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Cleanse the 数据总线 topic access cost raw data and aggregate it by topic dimension, excluding test topics, then write the results to a detail table.
Input: `internal_platform_db.ods_t_databus_access_topic_cost_date_d_query_engine_015`, partitioned by field `dt` (STRING, YYYYMMDD). Key fields: `s... | INSERT overwrite TABLE internal_platform_db.dwd_databus_topic_cost_detail_d_query_engine_015 PARTITION (dt = '20260507')
SELECT
MAX(systemname) AS system_belong,
MAX(dwproductname) AS category_name,
MAX(dwappgroup) AS dw_appgroup,
MAX(cityid) AS city_id,
MAX(iset) AS cluster_set,
topic,
MAX(data_size) AS ... | system_belong,category_name,dw_appgroup,city_id,cluster_set,topic,total_data_size_d,total_cost,in_charge,dt
SystemAlpha,数据仓库DWProdA,AppGroupX,101,ClusterA,topic_billing,2560000,162.0,user_zhang,20260507
SystemBeta,数据仓库DWProdB,AppGroupY,202,ClusterB,topic_log,1200000,81.0,user_li,20260507
SystemGamma,数据仓库DWProdC,AppGrou... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数(5) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,25,非key字段逐列值匹配率
D,D_field_completeness,field_completeness,15,关键字段非空/非空串比例
F,F_insert_overwrite,insert_overwrite,5... | tasks/offline-compute/HiveSQL/hivesql_006_en |
hivesql_007_en | offline-compute_HiveSQL_hivesql_007 | Aggregate Notebook instance runtime statistics by trace_id and p_date for the day: Input table w | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Aggregate Notebook instance runtime statistics grouped by `trace_id` and `p_date`.
Input: `internal_platform_db.dws_notebook_instance_execute_minute_stat_d_query_engine_019`
- Partition field: `dt` (STRING, format YYYYMMDD)
- Key fields: `trace_id`, `p_date`, `datawd_project_id`, `datawd_task... | INSERT overwrite TABLE internal_platform_db.dws_notebook_instance_execute_stat_d_query_engine_019 PARTITION (dt = '20260507')
SELECT
t1.trace_id,
t1.p_date,
t1.datawd_project_id,
t1.datawd_task_id,
t1.datawd_task_instance_id,
t1.compute_type,
t1.status_code,
t1.instance_run_time,
t1.code_run_time,
t... | trace_id,p_date,datawd_project_id,datawd_task_id,datawd_task_instance_id,compute_type,status_code,instance_run_time,code_run_time,resource_wait_time,code_start_time,code_end_time,instance_start_time,instance_end_time,serving_id,is_permanent,apply_for_gpu_count,code_gpu_count,instance_gpu_util,code_gpu_util,dt
trace_001... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数(5) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,25,非key字段逐列值匹配率
D,D_field_completeness,field_completeness,15,关键字段非空/非空串比例
F,F_insert_overwrite,insert_overwrite,5... | tasks/offline-compute/HiveSQL/hivesql_007_en |
hivesql_008_en | offline-compute_HiveSQL_hivesql_008 | Count tasks matching the shuffle_split tuning rule. Input table: internal_platform_db | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Count tasks that match the `shuffle_split` tuning rule, producing task-level tuning rule hit results.
Inputs:
- `internal_platform_db.nextgen_platform_dsl_spark_props_fht0_query_engine_021`
- `internal_platform_db.ods_spark_props_extra_query_engine_021`
- `internal_platform_db.deep_tuning_rul... | insert overwrite table internal_platform_db.shuffle_split_tuning_rule_hit_task_info_query_engine_021 partition (databus_imp_date = '20260507')
with spark_props as
(
select
t1.app_id
,max(spark_partition_min) spark_partition_min
,max(spark_partition_max) spark_partition_max
,max(round(ifnull(spark_ta... | usp_task_id,app_id,app_time_usage,tag,dw_appgroup,product_name,shuffle_split,max_shuffle_stage_task_num,shuffle_stage_time,shuffle_stage_task_time_50th,shuffle_stage_task_time_90th,spark_partition_min,spark_partition_max,spark_task_shuffle_size,spark_default_parallelism,spark_sql_shuffle_partitions,spark_shuffle_sort_b... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,20列 + 列名匹配
C,C_row_alignment,row_consistency,10,行数比例 + key覆盖率
D,D_shuffle_logic,shuffle_logic,20,shuffle_split 条件逻辑(CASE WHEN 版本判断)
D,D_deduplication,deduplication,15,ROW_NUMBER 按 usp_task_id 去重
D,D_numeric_values,numeric_valu... | tasks/offline-compute/HiveSQL/hivesql_008_en |
hivesql_009_en | offline-compute_HiveSQL_hivesql_009 | Compute GPU card-hours for task instances within 5-minute windows, joining Pod mapping and task instance configuration, outputting hourly instanc | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Compute GPU card-hours for task instances within 5-minute windows, outputting hourly instance-level GPU usage statistics.
Inputs:
- `internal_platform_db.gputj_gpu_info_parsed_query_engine_022`
- `internal_platform_db.dwd_ml_platform_instance_podname_query_engine_022`
- `internal_platform_db.... | INSERT OVERWRITE TABLE internal_platform_db.task_instance_gpu_time_stats_query_engine_022
PARTITION (dt='2026050700')
WITH
-- Step 1: 从GPU监控数据中提取pod运行记录,按分钟去重
pod_run_minutes AS (
SELECT
pod_name,
gpu_name,
FLOOR(CAST(pkg_time AS BIGINT) / 60) * 60 AS minute_timestamp,
FLOOR(CAST(pkg... | instance_uuid,time_5min,host_gpu_num,sum_run_time_m,gpu_hour,gpu_name,pod_count,host_num
inst-uuid-aaa,1767225600,8.0,7.0,0.9333333333333333,A100,2,2.0
inst-uuid-ccc,1767225600,2.0,1.0,0.03333333333333333,A100,1,1.0
inst-uuid-bbb,1767225900,4.0,2.0,0.13333333333333333,V100,1,1.0
inst-uuid-ddd,1767225900,8.0,1.0,0.13333... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,8列 + 列名匹配
C,C_row_alignment,row_consistency,10,行数比例 + key覆盖率
D,D_gpu_hour_calculation,gpu_hour_calculation,30,gpu_hour 核心指标计算正确性
D,D_window_aggregation,window_aggregation,15,"sum_run_time_m, pod_count 等窗口聚合指标"
D,D_gpu_config_l... | tasks/offline-compute/HiveSQL/hivesql_009_en |
hivesql_010 | offline-compute_HiveSQL_hivesql_010 | 从输入表中筛选状态为 killed 的 Notebook runner trace 数据,按天拆分后 | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
1) 任务目标
产出 killed 状态的 Notebook 实例 Pod 运行明细宽表。
2) 输入
- internal_platform_db.gputj_gpu_info_parsed_agg_1min_query_engine_024
- internal_platform_db.notebook_span_info_query_engine_024
- internal_platform_db.dwd_gputj_service_instance_map_query_engine_024
- internal_platform_db.dwd_ml_platform_instance_podname_... | INSERT OVERWRITE TABLE internal_platform_db.dwd_notebook_killed_instance_pod_detail_d_query_engine_024 PARTITION (dt = '20260507')
WITH base_trace AS (
SELECT trace_id
FROM internal_platform_db.notebook_span_info_query_engine_024
WHERE databus_imp_date >= '2026050700'
AND databus_imp_date <= '20260507... | trace_id,datawd_project_id,datawd_task_id,datawd_task_instance_id,compute_type,status_code,instance_run_time,code_run_time,resource_wait_time,code_start_time,code_end_time,instance_start_time,instance_end_time,serving_id,is_permanent,apply_for_gpu_count,pod_name,pkg_agg_time,gpu_util,gpu_count,p_date,dt
trace001,proj1,... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,22列 + 列名匹配
C,C_row_alignment,row_consistency,10,行数比例 + key覆盖率
D,D_time_calculation,time_calculation,25,"instance_run_time, code_run_time 等时间指标"
D,D_cross_day_split,cross_day_split,15,跨天 trace 按天拆分正确性
D,D_gpu_join,gpu_join,15,"... | tasks/offline-compute/HiveSQL/hivesql_010 |
hivesql_011 | offline-compute_HiveSQL_hivesql_011 | 从 notebook_span_info_query_engine_026 和 notebook_engin | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:从 Notebook span 信息和 engine 信息关联产出 killed 实例运行明细表。
输入:
- internal_platform_db.notebook_span_info_query_engine_026
- internal_platform_db.notebook_engine_info_query_engine_026
处理规则:
- 两表关联查询,关联条件需基于表结构确定
- 具体过滤条件和字段选择需基于表结构确定
输出要求:
- 输出表:internal_platform_db.dwd_notebook_killed_instance_detail_d_cand_qu... | INSERT overwrite TABLE internal_platform_db.dwd_notebook_killed_instance_detail_d_query_engine_026 PARTITION (dt = '20260507')
WITH base_trace AS (
SELECT trace_id
FROM internal_platform_db.notebook_span_info_query_engine_026
WHERE databus_imp_date >= '2026050700'
AND databus_imp_date <= '2026050700'
... | p_date,trace_id,datawd_project_id,datawd_task_id,datawd_task_instance_id,compute_type,status_code,instance_run_time,code_run_time,resource_wait_time,code_start_time,code_end_time,instance_start_time,instance_end_time,serving_id,is_permanent,apply_for_gpu_count,dt
2025-05-07,T001,P100,TASK001,INST001,ray,2,105,55,50,202... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,18列 + 列名匹配
C,C_row_alignment,row_consistency,10,行数比例 + key覆盖率
D,D_time_calculation,time_calculation,30,"instance_run_time, code_run_time 等时间指标"
D,D_cross_day_split,cross_day_split,15,跨天 trace 按天拆分正确性
D,D_engine_join,engine_joi... | tasks/offline-compute/HiveSQL/hivesql_011 |
hivesql_012_en | offline-compute_HiveSQL_hivesql_012 | From input table internal_platform_db.t_ed_socialbook_qq_high_v2_ta | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective
Filter out specified `tag_id` values from the tag incremental table and write the results to the target table.
Input
- `internal_platform_db.t_ed_socialbook_qq_high_v2_tag_guid_incr_query_engine_100`
Processing Rules
1. Filter condition: `ds = 20260608` AND `tag_id NOT IN (10212021124, 106060... | insert overwrite table internal_platform_db.t_ed_socialbook_qq_high_v2_tag_guid_incr_filter_query_engine_100 partition(ds=20260608)
select
guid,
tag_id,
tag_value,
redis_sub_key,
val_type,
splitter1,
splitter2,
limit_size,
tag_status
fro... | guid,tag_id,tag_value,redis_sub_key,val_type,splitter1,splitter2,limit_size,tag_status,ds
guid001,10212021100,val1,sub1,type1,:,",",10,1,20260608
guid005,10212021199,val5,sub5,type1,:,",",8,3,20260608 | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数(5) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,25,非key字段逐列值匹配率
D,D_field_completeness,field_completeness,15,关键字段非空/非空串比例
F,F_insert_overwrite,insert_overwrite,5... | tasks/offline-compute/HiveSQL/hivesql_012_en |
hivesql_013 | offline-compute_HiveSQL_hivesql_013 | 从 internal_platform_db.dws_ug_app_ad_material_actio | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:统计广告素材在各维度下的行为指标汇总,并生成一条更新日期记录。
输入:
- internal_platform_db.dws_ug_app_ad_material_action_di_query_engine_101
处理规则:
- 无 Join,单表处理
- 过滤条件:imp_date = 20260608
- 维度字段转换(使用 CASE WHEN):
- material_type: 1→'图片', 2→'视频'
- source: 2→'人工素材', 3→'人机素材', 6→'AI素材'
- video_model_type: 1→'镜头拆分模型(老模型)', 2→'剧情理解模型... | INSERT OVERWRITE TABLE internal_platform_db.ads_ug_app_ad_material_action_di_query_engine_101 PARTITION (imp_date=20260608)
SELECT
20260608 AS update_date
, CASE
WHEN material_type = 1 THEN '图片'
WHEN material_type = 2 THEN '视频'
... | update_date,material_type_name,source_name,video_model_type_name,create_type_name,purpose_name,provider_name,production_cnt,audit_cnt,pass_cnt,push_cnt,imp_cnt,clck_cnt,imp,clck,channel_name,imp_date
20260608,图片,人机素材,镜头拆分模型(老模型),原始剪辑,收入,AI侧,1,1,1,1,1,0,60,8,广告平台G,20260608
20260608,视频,人机素材,剧情理解模型(新模型),高光剪辑,拉活,内部平台D侧,1,1... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数17 + 列名匹配
C,C_row_alignment,row_consistency,15,行数比例 + key覆盖率
D,D_case_when_labels,case_when_labels,20,7个维度标签列逐行匹配
D,D_agg_metrics,agg_metrics,20,8个聚合列按权重匹配
F,F_union_all_sentinel,union_all_sentinel,10,是否存在NULL哨兵行
F,F_insert_... | tasks/offline-compute/HiveSQL/hivesql_013 |
hivesql_014_en | offline-compute_HiveSQL_hivesql_014 | Filter add-friend behavior from internal_platform_db.log_17047_query_engine_102 | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective:
Perform cluster aggregation analysis on add-friend behavior logs to identify malicious clusters, and output statistical metrics for each cluster.
Input:
- `internal_platform_db.log_17047_query_engine_102` (add-friend behavior log table)
Processing Rules:
1. No joins, single-table processing,... | INSERT INTO TABLE internal_platform_db.t_acct_addfri_action_cluster_minutely_query_engine_102 PARTITION(ds=202606090010)
select
appname_,
clientversion_,
scene_,
ticketscene_,
headmd5_,
uinipcountryid_,
uinipprovinceid_,
count(*) as addfri_pv,
count(distinct user_id_) as user_id_cnt,
sum(if(uinhighquality_=0, 1, 0)) as... | appname_,clientversion_,scene_,ticketscene_,headmd5_,uinipcountryid_,uinipprovinceid_,addfri_pv,user_id_cnt,low_quality_cnt,low_quality_rate,user_id_list,hello_content_list,evil_cnt,evil_rate,ds
app_hello_txt,800,1,100,headmd5_abc,86,440000,21,21,21,1.0,"1011,1010,1021,1020,1008,1019,1007,1018,1006,1017,1005,1016,1004,... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数16 + 列名匹配
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_agg_metrics,agg_metrics,40,"6个聚合指标列逐行匹配(addfri_pv,user_id_cnt,low_quality_rate,evil_rate,low_quality_cnt,evil_cnt)"
F,F_insert_overwrite,insert_overwrite... | tasks/offline-compute/HiveSQL/hivesql_014_en |
hivesql_015_en | offline-compute_HiveSQL_hivesql_015 | From internal_platform_db.t_mg_dws_user_tag_preset_241 | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Incrementally insert new user records (`user_id`, `loss_day`) into the `ds=20260608` partition of the target table that have not appeared before.
Inputs:
- `internal_platform_db.t_mg_dws_user_tag_preset_24128_query_engine_103` (source table, containing fields such as `ds`, `user_id`, `loss_da... | insert into internal_platform_db.t_mg_dws_user_tag_user_id_once_24128_query_engine_103 (user_id, loss_day)
select user_id,loss_day
from internal_platform_db.t_mg_dws_user_tag_preset_24128_query_engine_103
where
ds = 20260608
and user_id not in (select user_id from internal_platform_db.t_mg_dws_user_tag_user_id_once_24... | user_id,loss_day
user_a,5
user_b,3
user_e,1 | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数匹配 + 列名匹配
C,C_row_alignment,row_consistency,15,行数比例 + key覆盖率
D,D_anti_join,anti_join,25,NOT IN 排除已存在 user_id 的正确性
D,D_value_correctness,value_correctness,20,loss_day 等数值列逐行匹配
F,F_insert_overwrite,insert_mode,5,INSERT INTO(非 ... | tasks/offline-compute/HiveSQL/hivesql_015_en |
hivesql_016 | offline-compute_HiveSQL_hivesql_016 | 从用户活跃日志表统计每日各渠道的活跃用户数 | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:统计各渠道的日活跃用户数。
输入:
- internal_platform_db.dwd_user_activity_log_hi_query_engine_104(用户活跃日志表)
处理规则:
1. 过滤条件:imp_date = 20260608
2. 分组维度:channel
3. 聚合指标:COUNT(DISTINCT user_id) AS dau
4. 派生列:imp_date 固定值 20260608
输出要求:
- 字段顺序:imp_date(BIGINT,固定值20260608)、channel(STRING)、dau(BIGINT)
- 按 channel 升序排列
写入要求... | insert overwrite table internal_platform_db.dws_user_daily_active_by_channel_query_engine_104 partition(ds='20260608')
select
20260608 as imp_date
, channel
, count(distinct user_id) as dau
from
internal_platform_db.dwd_user_activity_log_hi_query_engine_104
where
imp_date = 20260608
group by
channel
order by
channel | imp_date,channel,dau,ds
20260608,appstore,3,20260608
20260608,huawei,2,20260608
20260608,oppo,1,20260608
20260608,xiaomi,1,20260608 | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数(4=3数据+1分区) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,25,非key字段逐列值匹配率
D,D_field_completeness,field_completeness,15,关键字段非空比例
F,F_insert_overwrite,insert_overwri... | tasks/offline-compute/HiveSQL/hivesql_016 |
hivesql_017 | offline-compute_HiveSQL_hivesql_017 | 从事件表和全量用户表 Join 过滤出 20260608 当天最后登录用户的事件,按 sdk_id | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:按 SDK 类型和小时统计活跃用户数。
输入:
- internal_platform_db.dws_event_tracking_ul2l6h5c_events_di_query_engine_114(事件明细表)
- internal_platform_db.dws_event_tracking_ul2l6h5c_all_user_df_query_engine_114(用户维度表)
处理规则:
- Join 条件:e.user_id = u.user_id
- 过滤条件:e.imp_date = 20260608,u.imp_date = 20260608,u.last_login_date ... | INSERT INTO internal_platform_db.ads_event_tracking_ul2l6h5c_user_active_hour_di_query_engine_114
SELECT
20260608 as `imp_date`
, ROW_NUMBER() OVER(ORDER BY e.`sdk_id`, e.`event_time_hour`) as `id`
, e.`sdk_id` as `sdk_id`
, e.`event_time_hour` as `hour`
, COUNT(DISTINCT e.user_id) as `active_users`
FROM
internal_platf... | imp_date,id,sdk_id,hour,active_users
20260608,1,sdk_android,10,2
20260608,2,sdk_ios,11,2 | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数(5) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,45,非key字段逐列值匹配率
F,F_insert_overwrite,insert_mode,5,INSERT INTO 写入模式
F,F_source_filter,source_coverage,5,2张源表均被引用
... | tasks/offline-compute/HiveSQL/hivesql_017 |
hivesql_018_en | offline-compute_HiveSQL_hivesql_018 | Compute scores for first-level comments from view records and comment logs. Group by recall_uin + channel_id + fe | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Compute the composite score for first-level comments newly added by recalled read users after their first post view, for content heat analysis.
Inputs:
- `internal_platform_db.dws_social_group_content_forum_hot_feed_recall_feed_view_hi_query_engine_122` (view records, filter `imp_hour` within... | insert overwrite table internal_platform_db.dws_social_group_content_forum_hot_feed_recall_comment_score_hi_query_engine_122 partition (imp_hour = 2026060910)
with social_group_feed_view as
(
select
uin as recall_uin
, channel_id
, feed_id
, min(first_view_time) as first_view_time
, max(last_view_time) as last_view_tim... | uin,channel_id,feed_id,comment_id,comment_time,comment_score,normal_comment_like_score,key_author_comment_like_score,key_owner_comment_like_score,normal_comment_reply_score,key_author_comment_reply_score,key_owner_comment_reply_score,avg_normal_comment_reply_score,normal_comment_like_uv,normal_comment_like_cnt,key_auth... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数26 + 列名匹配
C,C_row_alignment,row_consistency,10,"行数比例 + key(uin,channel_id,feed_id,comment_id)覆盖率"
D,D_comment_score_correctness,comment_score_correctness,25,comment_score + 7项分档评分列逐行匹配
D,D_uv_cnt_correctness,uv_cnt_correctne... | tasks/offline-compute/HiveSQL/hivesql_018_en |
hivesql_019_en | offline-compute_HiveSQL_hivesql_019 | From input table internal_platform_db.t_app_urlsafe_cont_topweb | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Extract qualifying potential risk domains from the top websites collection table and write them to the target table.
Input:
- `internal_platform_db.t_app_urlsafe_cont_topwebsites_collect_df_query_engine_125` (fields: `domain`, `site`, `ds`, `src`)
Processing Rules:
1. Filter condition: `src=... | INSERT OVERWRITE TABLE internal_platform_db.t_dws_urlsafe_cont_potential_risk_domain_di_query_engine_125
with topsites as
(
select domain as original_domain
from internal_platform_db.t_app_urlsafe_cont_topwebsites_collect_df_query_engine_125
where src='transco'
and ds=20260608
and length(split(domain, '\\.')[0])>3
limi... | ds,type,fuzzer,result_domain,original_domain
20260608,transco,-1,-1,long-domain-name.org
20260608,transco,-1,-1,test.site.net
20260608,transco,-1,-1,abcd.domain.cn | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,col_count(5) + col_names(5)
C,C_row_alignment,row_consistency,15,row_ratio(7) + key_coverage(8)
D,D_domain_filter_correct,domain_filter_correct,25,original_domain 列值匹配率(核心过滤逻辑)
D,D_fixed_value_correct,fixed_value_correct,20,ds... | tasks/offline-compute/HiveSQL/hivesql_019_en |
hivesql_020 | offline-compute_HiveSQL_hivesql_020 | 从 internal_platform_db.t_boss_v1_dict_item_hour_query_engine_130 筛选 | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:对指定分区的字典项小时表进行记录数监控统计,输出监控结果到目标表分区。
输入:
- internal_platform_db.t_boss_v1_dict_item_hour_query_engine_130
处理规则:
1. 无 Join,单表处理
2. 过滤条件:imp_hour=2026060910
3. 分组聚合:按 dim_name='AEGIS_ALL' 和 dim_value='AEGIS_ALL' 分组
4. 聚合计算:count(1) 得到 compute_item_51323
5. 派生列:
- imp_time = 2026060910
- data_type = ... | INSERT OVERWRITE TABLE internal_platform_db.t_boss_v1_dict_item_hour_monitor_res_query_engine_130 PARTITION (p_2026060910='p_2026060910', p_monitor='p_monitor')
SELECT imp_time, data_type, data_id, check_rule_id, check_item_id, dim_name, dim_value, compute_value, compare_value, check_value
FROM (
WITH temp_table_1 AS... | imp_time,data_type,data_id,check_rule_id,check_item_id,dim_name,dim_value,compute_value,compare_value,check_value,p_2026060910,p_monitor
2026060910,monitor,1::dept_om::t_boss_v1_dict_item_hour,51322,51323,AEGIS_ALL,AEGIS_ALL,3,,3,p_2026060910,p_monitor | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数(12=10数据+2分区) + 列名匹配(5)
C,C_row_alignment,row_consistency,15,行数比例(7) + key覆盖率(8)
D,D_field_value_match,field_value_match,45,非key字段逐列值匹配率
F,F_insert_overwrite,insert_overwrite,5,INSERT OVERWRITE + PARTITION 写入模式
F,F_partition... | tasks/offline-compute/HiveSQL/hivesql_020 |
hivesql_021_en | offline-compute_HiveSQL_hivesql_021 | Filter Chinese add-friend content from log table and write to hourly cluster table | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
### Task Objective
Filter Chinese content from the add-friend log table for the specified hour, deduplicate, and write to the hourly cluster table.
### Input
- `internal_platform_db.log_17047_query_engine_131`
### Processing Rules
1. No joins, single-table processing
2. Filter conditions:
- `day_ = '2026... | INSERT INTO TABLE internal_platform_db.data_team_member14_addcontent_cluster_hour_query_engine_131 PARTITION(ds='2026060914')
select distinct content_
from internal_platform_db.log_17047_query_engine_131
where day_='2026-06-09 00:00:00' and hour_ = '2026-06-09 14:00:00'
and commfrinum_=0
and appname_ in ('app_hello_txt... | content_
加好友一起聊天交流分享
你好我想加你好友认识
这是一个测试内容信息 | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数2(含分区列ds) + 列名匹配
C,C_row_alignment,row_consistency,15,行数比例(7) + key(content_)覆盖率(8)
D,D_values,values,40,content_列值集合匹配率
F,F_insert_overwrite,insert_overwrite,5,写入模式应为 INSERT INTO
F,F_partition_value,partition_value,5,ds='20... | tasks/offline-compute/HiveSQL/hivesql_021_en |
hivesql_022 | offline-compute_HiveSQL_hivesql_022 | 从 internal_platform_db.log_16159_query_engine_134 表中提取审核数据,过滤 ds=202 | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标:将日志表中符合特定条件的审核记录解析并转码,生成包含标签名称的可读审核结果表。
输入:
- internal_platform_db.log_16159_query_engine_134(包含审核日志字段:ds, functype_, actiontype_, othercol1_, resultinfo_, auditid_, operator_, strategyid_, orderid_, providerid_, uniqueauditid_, audittime_, createtime_, receivetime_, queue_label_)
处理规则:
1. 过滤条件:ds=2026... | insert overwrite table internal_platform_db.dwmid_daily_ecommerce_shop_level_func_2141_audit_result_query_engine_134 partition(ds ='2026060916')
with audit_result as
(
select 2026060916 ds,regexp_extract(othercol1_,'bizuin:(.*?);')bizuin,
regexp_extract(othercol1_,'nickname:(.*?);')nickname,
split(regexp_extract(result... | bizuin,nickname,account_tag1,account_tag2,account_tag3,account_tag1_string,account_tag2_string,account_tag3_string,remark,auditid_,operator_,strategyid_,orderid_,providerid_,uniqueauditid_,audittime,createtime_,receivetime_,queue_label_,ds
biz002,shop_B,8001,6001,5002,商品品类杂糅,引人不适(严重),违法违禁,another_remark,5002,op_user2,1... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,10,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数20 + 列名匹配
C,C_row_alignment,row_consistency,10,"行数比例 + key(uniqueauditid_,strategyid_,orderid_,providerid_)覆盖率"
D,D_regex_extraction,regex_extraction,20,"正则提取字段逐行匹配(bizuin,nickname,account_tag1/2/3,remark)"
D,D_account_tag_m... | tasks/offline-compute/HiveSQL/hivesql_022 |
hivesql_023_en | offline-compute_HiveSQL_hivesql_023 | Compute multi-dimensional CUBE aggregation of user open-start-path data by referer_type, referer, and path | offline-compute | HiveSQL | offline-compute/HiveSQL | en | pure-text | 600 | ## Prompt
Task Objective: Compute multi-dimensional aggregation metrics from the user open-start-path detail table, outputting UV and first-time UV across three dimension combinations: application category (`referer_type`), source application (`referer`), and launch path (`path`).
Inputs:
- `internal_platform_db.t_ed_... | insert overwrite table internal_platform_db.t_md_mapservice_user_open_start_path_dwa_di_query_engine_135 partition(ds=20260608)
with t as (
select
event_time,platform,app_version,channel,t1.uin,event_code,event_value,city,brand,device_id_type,bg,event_timestamp,referer,path,is_first
,case when referer in ('com.food... | referer_type,referer,path,uv,first_uv,ds
FoodChainA,com.foodchain_a.android.activity,qqmap://map/routeplan?type=walk,1,1,20260608
CourierCoB,couriercob,qqmap://map/navi,1,0,20260608
CourierCoB,total,total,2,0,20260608
CourierCoB,total,qqmap://map/search,1,0,20260608
CourierCoB,total,qqmap://map/navi,1,0,20260608
Courie... | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数6 + 列名匹配
C,C_row_alignment,row_consistency,15,"行数比例 + key(referer_type,referer,path)覆盖率"
D,D_path_clean_referer_map,path_clean_referer_map,25,"路径清洗+referer映射维度值逐行匹配(referer_type,referer,path)"
D,D_cube_aggregation,cube_aggre... | tasks/offline-compute/HiveSQL/hivesql_023_en |
hivesql_024 | offline-compute_HiveSQL_hivesql_024 | 从 internal_platform_db.dwd_relationship_strength_fe | offline-compute | HiveSQL | offline-compute/HiveSQL | zh | pure-text | 600 | ## Prompt
任务目标
基于好友关系特征数据计算关系强度推荐分数,输出每个用户对(uin, touin)的加权评分结果。
输入
- internal_platform_db.dwd_relationship_strength_features_v4_di_query_engine_138(无 Join,单表处理)
处理规则
1. 过滤条件:imp_date = 20260608,且 uin >= 10000 且 touin >= 10000
2. 特征处理:
- 对所有分数字段使用 COALESCE(field, 0) 填充空值
- c2c_cnt_score、common_frd_num、frd_tag_nu... | INSERT OVERWRITE TABLE internal_platform_db.ads_qq_sq_frd_recommendation_result_list_df_query_engine_138 PARTITION (imp_date = 20260608)
WITH feature_modified_v2 AS (
SELECT
uin
, touin
, COALESCE(c2c_score, 0) AS c2c_score
, COALESCE(LOG(1 + c2c_cnt_score), 0) AS log_c2c_cnt_score
, COALESCE(socialzone_visit_score, 0... | uin,touin,score,raw_score
10009,10010,570.0,0.0569
10001,10002,307.0,0.0306
10003,10004,238.0,0.0237
10007,10008,31.0,0.003 | 维度,维度全名,子维度,满分,说明
A,A_executability,executability,15,result.sql 能跑通且产出非空
B,B_schema,schema,10,列数5 + 列名匹配
C,C_row_alignment,row_consistency,15,"行数比例 + key(uin,touin)覆盖率"
D,D_score_correctness,score_correctness,25,最终得分列逐行匹配
D,D_raw_score_correctness,raw_score_correctness,20,原始分列逐行匹配
F,F_null_handling,null_handling,5,COAL... | tasks/offline-compute/HiveSQL/hivesql_024 |
DataClawEval
An executable benchmark for end-to-end data-engineering agents in industrial environments.
DataClawEval measures an autonomous agent's ability to inspect data, implement and debug pipelines, and materialize correct artifacts in realistic data-engineering workflows. It contains 100 production-grounded tasks across five execution engines: PySpark, MySQL, HiveSQL, PrestoSQL/Trino, and FlinkSQL. Each task runs in an isolated Docker sandbox and is evaluated by a case-specific, deterministic, rule-based grader.
What's Inside
- Production-grounded tasks — 100 end-to-end tasks reconstructed from production-grade implementations written by professional enterprise data engineers and validated through execution checks, expert perturbations, and expert review.
- Batch and streaming coverage — batch transformations in PySpark, MySQL, HiveSQL, and PrestoSQL/Trino, plus streaming-oriented FlinkSQL tasks involving event time, watermarks, windows, joins, and aggregations.
- End-to-end agent workflows — schema inspection, data exploration, implementation, execution, debugging, output validation, and artifact materialization.
- Artifact- and process-oriented grading — task-specific graders assess executable outputs and engineering behaviors such as exploration, execution efficiency, and self-verification.
- Reproducible environments — every run starts in a fresh container with task-specific data, services, execution engines, and deterministic initialization.
- Flexible agent integration — run evaluations with the
codebuddy,claude-code, orcodexbackend. The paper evaluates 16 model configurations through a unified CodeBuddy scaffold. - Batch experiments — evaluate one or more engines in parallel and generate combined reports.
Task Suite
| Execution engine | Workload | Tasks |
|---|---|---|
| PySpark | Batch (offline-compute) |
20 |
| MySQL | Batch (offline-compute) |
20 |
| HiveSQL | Batch (offline-compute) |
28 |
| PrestoSQL/Trino | Batch (offline-compute) |
12 |
| FlinkSQL | Streaming (online-compute) |
20 |
| Total | 100 |
The suite is also balanced between 50 English and 50 Chinese task prompts and covers five business domains: Ops & Resource Governance (30), Data Analytics & User Growth (30), Security & Risk Control (16), Content, Community & Dev-Efficiency (14), and Advertising & Marketing (10).
Each task lives in tasks/<workload>/<engine>/<task_id>/ and contains:
task.md— the natural-language request and metadata such as timeout and engineinit/— initialization or verification assets; batch tasks initialize tables here, while FlinkSQL tasks generally define generated streaming inputs in the submitted SQLgt/— the reference implementation and task-specificgrade.pyused during evaluation
How It Works
For every model–task run, the harness:
- Starts a fresh Docker container from the configured
DOCKER_IMAGE. - Starts the services required by the selected engine and initializes task-specific inputs.
- Prepares the agent workspace with the task prompt and engineering tools.
- Runs the selected agent backend and records its execution trajectory.
- Loads
gt/after agent execution and runs the task-specific grader in the same container. - Collects the submitted solution files, workspace snapshot, scores, usage data, and execution
transcript into
output/, then removes the container.
Requirements
| Component | Requirement |
|---|---|
| Docker | Docker CLI and a running daemon |
| Python | Python 3.11 on the host; the scripts invoke python3.11 |
| Storage | Enough free space for the multi-engine image and task outputs |
| Model access | Credentials and network access for the selected agent backend |
The evaluation image provides a ready-to-use multi-engine environment with Spark 3.5.0, Flink 1.18.1, Trino 435, MySQL, a Hive Metastore, Python 3.11.13, and the supported agent runtimes.
Quick Start
0. Verify prerequisites
Make sure Docker is installed and running, and that Python 3.11 is available.
docker version # verifies the Docker CLI and daemon connection
python3.11 --version # Python 3.11 is required
1. Install Python dependencies
python3.11 -m pip install -r requirements.txt
2. Get the evaluation image
You have two options.
Option A — Pull the prebuilt image (recommended)
Skip the local build and pull the ready-made image from the registry:
docker pull dicemy/dataclaweval:v1.0
docker tag dicemy/dataclaweval:v1.0 dataclaw-eval:v1.0 # match DOCKER_IMAGE in .env
The default
DOCKER_IMAGEin.envisdataclaw-eval:v1.0. Either retag the pulled image as shown above, or setDOCKER_IMAGE=dicemy/dataclaweval:v1.0in your.env.
Option B — Build locally
prepare.sh vendors the installed CodeBuddy Agent SDK and builds the
dataclaw-eval:v1.0 evaluation image.
bash script/prepare.sh
3. Configure credentials
Copy the example env file and fill in the values for the agent backend you plan to use.
cp .env.example .env
Then edit .env:
DOCKER_IMAGE=dataclaw-eval:v1.0
# CodeBuddy auth (required for the `codebuddy` harness)
CODEBUDDY_AUTH_TOKEN=
CODEBUDDY_API_KEY=
CODEBUDDY_INTERNET_ENVIRONMENT=ioa # public | internal | ioa
# OpenRouter (required for `claude-code` and `codex` harnesses)
OPENROUTER_API_KEY=
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
Get a CodeBuddy API key: overseas → https://www.codebuddy.ai/profile/keys · China → https://copilot.tencent.com/profile/ · iOA (Tencent staff) → https://tencent.sso.copilot.tencent.com/profile/keys. Set
CODEBUDDY_INTERNET_ENVIRONMENTto match your account type.
4. Run the evaluation
# Run a single execution engine
bash script/run_batch.sh -P # PySpark
bash script/run_batch.sh -M # MySQL
bash script/run_batch.sh -H # HiveSQL
bash script/run_batch.sh -F # FlinkSQL
bash script/run_batch.sh -PR # PrestoSQL/Trino
# Run workload groups
bash script/run_batch.sh -offline # PySpark + MySQL + HiveSQL + PrestoSQL/Trino
bash script/run_batch.sh -online # FlinkSQL
bash script/run_batch.sh -all # all 100 tasks
Use script/run_batch.sh -all to evaluate the complete 100-task suite. Results and a combined
report are written to output/.
Usage Reference
run_batch.sh options:
| Flag / Option | Description |
|---|---|
-P -M -H -F -PR |
Select PySpark / MySQL / HiveSQL / FlinkSQL / PrestoSQL/Trino |
-offline -online -all |
Run predefined workload groups |
--parallel N |
Number of tasks to run concurrently (default: 10) |
--harness NAME |
Agent backend: codebuddy (default), claude-code, codex |
--model MODEL |
Model name (repeatable, codebuddy only) for multi-model runs |
Examples:
# PySpark with 5 workers
bash script/run_batch.sh -P --parallel 5
# Use the Claude Code harness
bash script/run_batch.sh -P --harness claude-code
# Compare two CodeBuddy models on PySpark
bash script/run_batch.sh -P --model model_a --model model_b
To run a single task directly (bypassing batch mode):
python3.11 eval/run_batch.py \
--agent-backend codebuddy \
--task tasks/offline-compute/PySpark/pyspark_001/task.md \
--model deepseek-v4-flash-ioa
Output & Scoring
Each run is saved under output/<harness>/<category>/<task_id>/<run_suffix>/ with the submitted
solution, workspace snapshot, transcript, usage statistics, and score.json.
DataClawEval evaluates both the final data product and the engineering workflow:
- Artifact quality — executability, schema correctness, row-level alignment, numerical accuracy, and categorical or business correctness.
- Process quality — exploration adequacy, execution efficiency, and self-verification.
The overall score combines both dimensions:
[ S = \alpha S_{artifact} + (1-\alpha)S_{process} ]
The artifact weight (\alpha) is configured per task, with 0.7 as the most common value. Repository
outputs normalize overall_score to 0–1, while the paper presents scores on a 0–100 scale.
Batch runs aggregate model and engine summaries into
output/experiment_report_<timestamp>.json and report the average score and the number of tasks
with overall_score >= 0.5:
Category/Model Tasks Avg Score Pass(>=0.5)
offline-compute/PySpark/model_a 20 0.7350 15/20
Paper Results
The paper evaluates 16 model configurations from eight families on all 100 tasks: 1,600 primary runs in total. Every model uses the same fixed Tencent CodeBuddy scaffold, and only the underlying LLM changes. The primary table uses one run per model–task pair; task prompts are evenly split between English and Chinese (50 each).
| Model | Overall score (0–100) | Avg. tokens/task |
|---|---|---|
| GPT 5.5 | 74.9 | 299.8k |
| Claude Opus 4.8 | 74.3 | 318.3k |
| Claude Sonnet 5 | 73.8 | 457.9k |
| Gemini 3.1 Pro | 73.7 | 292.4k |
| Gemini 3.5 Flash | 73.3 | 973.8k |
| DeepSeek V4 Flash | 73.0 | 419.6k |
| MiniMax M3 | 71.8 | 714.1k |
| GLM 5.1 | 71.6 | 355.4k |
| DeepSeek V4 Pro | 70.6 | 359.3k |
| Kimi K2.6 | 69.0 | 428.8k |
| GLM 5.2 | 68.8 | 403.9k |
| Kimi K2.7 | 68.1 | 407.5k |
| GPT 5.3 Codex | 66.4 | 271.4k |
| Hy3 | 66.0 | 468.7k |
| MiniMax M2.7 | 63.7 | 501.0k |
| GLM 5V Turbo | 60.3 | 287.8k |
Research highlights:
- GPT 5.5 achieves the highest overall score at 74.9.
- Engine-level results reveal complementary strengths: Claude Opus 4.8 leads PySpark; GPT 5.5 leads HiveSQL and MySQL; DeepSeek V4 Pro and Gemini 3.5 Flash tie on PrestoSQL/Trino; and DeepSeek V4 Flash leads FlinkSQL.
- Token and tool-call analyses identify models that combine strong task performance with efficient agent execution.
- Repeated-run experiments add execution consistency as a first-class evaluation dimension.
- Execution-grounded, rule-based graders provide stable, case-specific assessment of both generated artifacts and engineering processes.
See the paper for complete engine-level results, efficiency and stability analyses, bilingual performance, and all task listings.
Project Structure
DataClawEval/
├── Dockerfile # Self-contained sandbox (Spark, Flink, Trino, MySQL, Python 3.11)
├── requirements.txt # Host-side Python dependencies
├── .env.example # Credentials template
├── eval/
│ └── run_batch.py # Evaluation entry point
├── script/
│ ├── prepare.sh # Vendor SDK + build the Docker image
│ ├── run_batch.sh # Batch runner + report generator
│ ├── run.sh # Thin wrapper around run_batch.py
│ └── ... # Metastore / Trino / entrypoint helpers
├── src/
│ ├── agents/ # Agent backends: codebuddy, claudecode, codex, common
│ └── utils/ # Docker orchestration, DB init, grading, task parsing
└── tasks/
├── offline-compute/ # PySpark, MySQL, HiveSQL, PrestoSQL
└── online-compute/ # FlinkSQL
Citation
If you use DataClawEval in research, cite the project using CITATION.cff. The
archived software release is available at https://doi.org/10.5281/zenodo.21621566.
License
Released under the MIT License.
- Downloads last month
- 3