Four CN Lite tasks have missing or mismatched source data (tasks 33, 207, 380, and 381)

[eissac]

Summary

Four tasks in the Chinese Workspace-Bench Lite dataset appear to have missing or mismatched source data:

• opendatabox-workspace-bench-33
• opendatabox-workspace-bench-207
• opendatabox-workspace-bench-380
• opendatabox-workspace-bench-381

I checked:

1. The task metadata, data_manifest, and file_dep_graph
2. The actual files distributed with each Lite task
3. The complete filename index of filesys_cn.zip from Workspace-Bench/Workspace-Bench-Workspaces (28,023 entries)
4. The provided rubric-grounding audit

The issues are described below.

---

1. opendatabox-workspace-bench-33: hospital-grade data is unavailable

The task asks the agent to calculate the numbers of level-1, level-2, and level-3 hospitals in eastern, central, and western China.

The supplied inputs are:

• 1-12 2023年各地区按床位数分组的社区卫生服务中心(站)数.xlsx
• 1-2 2023年各地区医疗卫生机构数.xlsx
• 1-3 2023年各类医疗卫生机构数.xlsx

The files contain:

• Regional and province-level hospital totals
• Hospital counts by institution type
• Community health center/station bed-count groups
• National institution classifications

They do not contain regional or province-level hospital-grade fields.

A search of the complete Operations Manager workspace found no hospital-grade distribution workbook. The closest file is:

• 医疗分析/医院收支/4-12 2023年三级公立医院收入与支出.xlsx

That workbook contains national financial indicators by hospital grade, not regional or province-level hospital-grade counts.

As a result:

• The eastern, central, and western hospital totals are grounded.
• The community health center/station statistics are grounded.
• The required level-1, level-2, and level-3 counts for the three regions are not independently derivable.
• The Beijing and Shanghai hospital totals are grounded, but their grade-specific counts are not.

The current audit marks 10 of 19 rubrics as grounded and 9 as ungrounded.

Suggested maintainer action

Either:

1. Add the intended regional hospital-grade source workbook to the Operations Manager workspace and the task manifest; or
2. Remove the hospital-grade calculations and regenerate the affected rubrics using the available institution-type and bed-group data.

---

2. opendatabox-workspace-bench-207: required scoring-model workbook is missing

The task explicitly says:

基于5-通用人才画像（模型）.xlsx评价四份简历，并且生成人才评价.xlsx到桌面

However, the task manifest contains only four resume files:

• 张浩然简历.docx
• 王佳宁简历.docx
• 赵思远简历.docx
• 李雨辰简历.docx

The required 5-通用人才画像（模型）.xlsx is absent from both the task manifest and file_dep_graph.

A full search of the Logistics Manager workspace found no file named:

• 5-通用人才画像（模型）.xlsx
• 5-通用人才画像(模型).xlsx

It also found no filename containing 通用人才画像.

The workspace contains only these related templates:

• 人才画像/模板/2-人才画像矩阵图.xlsx
• 人才画像/模板/3-人才画像_履历模型_.xlsx
• 人才画像/模板/4-人才画像_冰山模型_.xlsx
• 人才画像/模板/6-核心岗位人才画像_模版_.docx

None can safely be assumed to be the missing model.

Without the model workbook, an agent cannot independently recover:

• The five scoring dimensions and their exact weights
• The grade thresholds
• The intended scoring formulas
• Exact candidate scores such as 82, 85, 88, and 92

Resume-derived facts remain grounded, but the scoring-model results do not. The current audit marks 10 of 20 rubrics as grounded and 10 as
ungrounded.

Suggested maintainer action

Recover and add 5-通用人才画像（模型）.xlsx to:

• The Logistics Manager workspace
• data_manifest
• file_dep_graph
• Any generated setup/input manifest

If the original workbook contains expected candidate answers, please publish a sanitized template containing only task-visible criteria,
weights, formulas, and grade thresholds, and regenerate rubrics that depend on hidden scores.

---

3. opendatabox-workspace-bench-380: prompt and source assets describe different business scenarios

The prompt claims the following assets are provided:

• activity_photos/
• impact_chart.csv
• feedback_wordcloud.png
• future_plan.md
• recruitment_poster.png

None of these paths exists in the complete Chinese workspace image.

The actual task manifest contains:

• post_1.json
• market_analysis_1.json
• user_feedback_category_202601.md
• strategic_plan_1.md
• training_module_1.md

These files describe corporate marketing operations rather than a volunteer association:

• post_1.json is an Instagram marketing post and contains only a remote example.com image URL.
• market_analysis_1.json describes market size, growth, and competitors.
• user_feedback_category_202601.md contains product, UX, and customer-service feedback.
• strategic_plan_1.md contains customer-growth, product, and revenue targets.
• training_module_1.md is a Marketing Operations training course.

Importantly, the files are not unrelated to the rubrics. Most exact rubric values come directly from these actual marketing files, including:

• Reach of 87,170,000
• 2,070,100 new followers
• A 3.51% engagement rate
• Market size 318990B, growth 6%, and CAGR 11%
• Feedback counts and percentages
• The 2024–2026 revenue targets
• The four-hour online Marketing Operations course

The current audit therefore marks 21 of 25 rubrics as grounded.

The defect is a semantic mismatch among:

• A volunteer-association prompt
• Corporate marketing inputs
• Rubrics primarily generated from the marketing inputs

Renaming the current files would not resolve the mismatch: a marketing post is not an activity-photo collection, and a marketing training
module is not a recruitment poster.

Suggested maintainer action

Either:

1. Add the five intended volunteer-association assets and regenerate all content rubrics from them; or
2. Rewrite the prompt as a corporate marketing annual-review task, retaining the current inputs and grounded rubric values while removing the
   unsupported volunteer, activity-photo, word-cloud, and recruitment-poster requirements.

Given the existing rubric content, the second option may require fewer changes.

---

4. opendatabox-workspace-bench-381: filename, year, format, data-grain, and scoring-model mismatch

The prompt names five 2025 CSV files:

• hospital_finance_2025.csv
• liabilities_2025.csv
• drug_production_2025.csv
• medical_expenses_2025.csv
• chronic_disease_2025.csv

None exists in the complete Chinese workspace image.

The actual task inputs are five Chinese-named 2023 XLSX workbooks:

• 4-12 2023年三级公立医院收入与支出.xlsx
• 4-6 2023年各类医疗卫生机构资产与负债.xlsx
• 4-4-2 2023年全国药品生产流通.xlsx
• 4-20 2023年各地区公立医院门诊和住院病人次均医药费用.xlsx
• 9-10 2023年调查地区15岁及以上居民慢性病患病率.xlsx

This is not only a filename mismatch. The data grains are incompatible with the requested province-level correlation analysis:

• Medical expenses: 31 province-level regions
• Drug production/distribution: 31 province-level regions
• Chronic disease: national, urban/rural, and broad eastern/central/western aggregates only
• Hospital income and expenditure: hospital-grade aggregates, not provinces
• Assets and liabilities: institution-type aggregates, not provinces

Therefore, the inputs cannot be merged into a 31-region dataset containing both financial-health and chronic-disease variables.

The task also does not define:

• The financial-health score
• The balance/surplus ratio formula
• Risk-score weights
• Low/medium/high-risk thresholds
• How the exact 5/23/3 risk split should be produced

There are also apparent rubric inconsistencies:

• A value of 86.06% appears more consistent with an expense-to-income ratio than a conventional surplus ratio. A conventional surplus ratio
  would instead be approximately 13.94%.
• If drug-production licenses are used as the industry-scale variable, the correlation with outpatient cost is approximately r = +0.108,
  not negative.
• Using drug-distribution-business licenses instead gives approximately r = -0.137, showing that the required negative conclusion depends
  on an unspecified metric choice.

The current audit marks only 3 of 20 rubrics as grounded and 17 as ungrounded.

Suggested maintainer action

This task likely needs to be regenerated rather than fixed through filename aliases alone.

Either:

1. Add the intended 2025 province-level CSV datasets, explicitly define all formulas and risk thresholds, and regenerate the rubrics; or
2. Rewrite the task around the available 2023 XLSX workbooks, restrict analysis to supported aggregation levels, define all derived metrics,
   and regenerate the correlation and risk rubrics.

---

Expected outcome

Please either correct the affected source assets and manifests or revise the task contracts and regenerate rubrics so that every required
result can be independently derived from task-visible inputs.

[wu-uk]

English

Thank you for the systematic audit of data derivability. An opt-in local patch is available in PR #28: #28. To preserve fair comparison with existing leaderboard submissions, we are not changing the currently published Hugging Face snapshot at this time. Users who need the corrected task contracts can explicitly run the patch command below. We also plan to release a new Workspace-Bench version with refreshed task metadata and assets soon.

We addressed the four tasks using the principle that every required result must be independently obtainable from task-visible inputs.

• Task 33
  • The supplied inputs do not contain regional hospital-grade distributions, so the task no longer requires level-1, level-2, or level-3 hospital counts.
  • It now covers source-supported hospital totals, hospital-type distributions, and community health service center/station bed configurations.
  • The output filename and rubrics were updated accordingly, and the task explicitly says not to infer hospital grades from hospital types.
• Task 207
  • Added a sanitized 5-通用人才画像（模型）.xlsx containing the five evaluation dimensions, weights, score bands, and general scoring rules.
  • Added the model workbook to data_manifest and file_dep_graph, with a target_path so it is materialized into the runtime workspace.
  • Removed unsupported Putonghua sub-level requirements and fixed candidate-score expectations that depended on hidden answer anchors rather than visible resume evidence.
• Task 380
  • Rewrote the task as a marketing-operations annual-review task, consistent with the five supplied assets and the grounded content requirements.
  • Changed the output to Marketing_Annual_Review.pptx and the destination to /营销运营/.
  • Removed unsupported volunteer-association, activity-photo, word-cloud, and recruitment-poster assumptions.
• Task 381
  • Removed the unsupported province-level financial-health/chronic-disease risk score.
  • The revised task analyzes hospital income and expenses by hospital grade, assets and liabilities by institution type, regional medical expenses, regional drug-license counts, and national/regional chronic-disease prevalence at their actual aggregation levels.
  • It explicitly disallows forced cross-granularity merges. The surplus-rate definition is corrected to (total income - total expenses) / total income, which is approximately 4.99% for public hospitals nationally.
  • The correlation rubric now declares the metric choice: when drug-production license counts are used, the Pearson correlation with outpatient per-visit medical expense is approximately +0.108. The unsupported negative-correlation claim and 5/23/3 risk split were removed.

These corrections are provided as an explicit local patch and are not applied automatically after download. The published Hugging Face snapshot remains unchanged for leaderboard compatibility. After downloading the unchanged snapshot, users who need the corrected contracts can run:

cd evaluation
python3 scripts/apply_task_patches.py --kind lite --language cn

Test coverage is included.

中文

感谢对数据可推导性的系统性审计。可选的本地 patch 已提交至 PR #28：https://github.com/OpenDataBox/Workspace-Bench/pull/28。为了与现有 leaderboard 结果公平对比，我们暂不修改当前已发布的 Hugging Face 快照；如需使用修正后的任务合同，可显式运行下方 patch 命令。我们也计划很快发布包含更新后任务元数据和资源的新版本 Workspace-Bench。

我们按照“每一项必需结果都应能由任务可见输入独立推导”的原则处理了这 4 个任务。

• Task 33
  • 当前输入不包含地区医院等级分布，因此任务不再要求一级、二级、三级医院数量。
  • 新任务改为覆盖有来源支持的医院总数、医院类型分布和社区卫生服务中心（站）床位配置。
  • 输出文件名和 rubrics 已同步调整，且任务明确禁止从医院类型推断医院等级。
• Task 207
  • 已补充脱敏的 5-通用人才画像（模型）.xlsx，包含五个评价维度、权重、评分区间和通用评分规则。
  • 该模型工作簿已加入 data_manifest 和 file_dep_graph，并通过 target_path 在运行时 materialize 到 workspace。
  • 已移除输入简历无法支持的普通话“甲/乙等”细分要求，以及依赖隐藏答案锚点的固定候选人分数要求。
• Task 380
  • 已将任务重写为与 5 份现有输入和有依据内容要求一致的营销运营年度复盘任务。
  • 输出改为 Marketing_Annual_Review.pptx，目标目录改为 /营销运营/。
  • 已移除志愿者协会、活动照片、词云和招新海报等无输入支持的假设。
• Task 381
  • 已移除现有数据无法支持的省级“财务健康度—慢性病”风险评分。
  • 修订后的任务按真实聚合层级分析医院等级收支、机构类型资产负债、地区医药费用、地区药品许可证数量，以及全国/区域慢性病患病率。
  • 任务明确禁止强行跨粒度合并。结余率口径已修正为 (总收入 - 总费用) / 总收入，全国公立医院约为 4.99%。
  • 相关性 rubric 现已明确指标口径：使用药品生产许可证数量时，其与门诊次均医药费用的 Pearson 相关系数约为 +0.108；原先无依据的负相关结论和 5/23/3 风险划分已移除。

这些修复以显式本地 patch 的形式提供，下载后不会自动应用。为保持 leaderboard 可比性，当前 Hugging Face 快照维持不变。下载未修改的快照后，如需使用修正后的任务合同，可执行：

cd evaluation
python3 scripts/apply_task_patches.py --kind lite --language cn

相关测试覆盖已补充。