Rename dataset branding from DataClaw to DataClawBench in READMEs
Browse files- .gitattributes +1 -0
- README.md +2 -2
- README.zh-CN.md +6 -6
.gitattributes
CHANGED
|
@@ -59,3 +59,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
assets/database/**/*.csv filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
assets/database/**/*.csv filter=lfs diff=lfs merge=lfs -text
|
| 62 |
+
assets/database_clear/**/*.csv filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -57,7 +57,7 @@ DataClawBench is designed for exactly this shift. It evaluates not abstract capa
|
|
| 57 |
|
| 58 |
DataClawBench is a process-oriented data-analysis benchmark for realistic, complex data environments. Its core goal is not merely to measure agents' end-task performance, but to serve as a high-fidelity testbed that also evaluates, at fine granularity, how agents evolve when facing real-world complexity and multi-step reasoning.
|
| 59 |
|
| 60 |
-
DataClawBench simulates at scale the noisy, weakly-semantic, cross-domain data environments found in the real world. Complex data-analysis questions are authored by domain experts in finance and computer science, and each task's process annotations and unique objective answers are cross-verified by human experts with AI assistance. Process annotations include task milestones, human-corrected reference trajectories, and evidence data sources.
|
| 61 |
|
| 62 |
|
| 63 |
## 🎯 Why DataClawBench?
|
|
@@ -360,7 +360,7 @@ Under a theme-domain view and business-oriented taxonomy, the current data envir
|
|
| 360 |
</tbody>
|
| 361 |
</table>
|
| 362 |
|
| 363 |
-
> At execution time, agents typically need to align entities across files, join across tables, normalize definitions, and perform aggregation calculations, rather than simply looking up values in a single file; when needed, they must also consult business conventions in `internal_metrics.csv`. This is the core value of
|
| 364 |
|
| 365 |
### 📋 Task Statistics
|
| 366 |
|
|
|
|
| 57 |
|
| 58 |
DataClawBench is a process-oriented data-analysis benchmark for realistic, complex data environments. Its core goal is not merely to measure agents' end-task performance, but to serve as a high-fidelity testbed that also evaluates, at fine granularity, how agents evolve when facing real-world complexity and multi-step reasoning.
|
| 59 |
|
| 60 |
+
DataClawBench simulates at scale the noisy, weakly-semantic, cross-domain data environments found in the real world. Complex data-analysis questions are authored by domain experts in finance and computer science, and each task's process annotations and unique objective answers are cross-verified by human experts with AI assistance. Process annotations include task milestones, human-corrected reference trajectories, and evidence data sources. DataClawBench adopts OpenClaw as its unified agent framework.
|
| 61 |
|
| 62 |
|
| 63 |
## 🎯 Why DataClawBench?
|
|
|
|
| 360 |
</tbody>
|
| 361 |
</table>
|
| 362 |
|
| 363 |
+
> At execution time, agents typically need to align entities across files, join across tables, normalize definitions, and perform aggregation calculations, rather than simply looking up values in a single file; when needed, they must also consult business conventions in `internal_metrics.csv`. This is the core value of DataClawBench for evaluating real-scenario data understanding and reasoning capabilities.
|
| 364 |
|
| 365 |
### 📋 Task Statistics
|
| 366 |
|
README.zh-CN.md
CHANGED
|
@@ -29,18 +29,18 @@
|
|
| 29 |
|
| 30 |
DataClawBench 正是面向这一变化而设计。它评测的不是脱离执行环境的抽象能力,而是在真实数据条件、明确任务约束和可复现执行协议下,OpenClaw 类端到端智能体完成数据分析任务的实际表现。
|
| 31 |
|
| 32 |
-
## 🔍
|
| 33 |
|
| 34 |
DataClawBench 是一个面向真实复杂数据环境的过程导向数据分析任务评测基准。核心目标并非仅仅是衡量智能体在数据分析任务上的最终表现,而是旨在将其作为一个现实高保真的测试平台,同时细粒度地评估智能体在面临真实复杂环境和多步推理时的演化过程。
|
| 35 |
|
| 36 |
-
DataClawBench 大规模地模拟了高噪声、弱语义、跨领域的真实世界数据环境。通过金融领域与计算机领域的人类专家撰写复杂数据分析任务问题,并由人类专家和AI辅助交叉核验提供每个任务的过程性标注和唯一客观答案。其中,过程性标注包括任务里程碑、人工订正参考轨迹和证据数据来源。
|
| 37 |
|
| 38 |
|
| 39 |
## 🎯 为什么选择 DataClawBench
|
| 40 |
|
| 41 |
- **从理想化的假设数据环境到非理想化的真实数据环境。** DataClawBench 包含混合结构和非结构化数据,覆盖企业画像、企业经营状态、区域产业统计、全国行业统计与政策文本等资源。所有信息均来自真实世界采集,并面临指标缺失、口径不一致、命名不一致等摩擦;任务面对的是现实风格的数据环境,而非过度清洗后的单表查值。
|
| 42 |
-
- **从单点静态查询到多步动态推理。**
|
| 43 |
-
- **从结果导向评估到过程导向评估。** DataClawBench超越单纯的结果准确率评估,对智能体的运行演化过程进行解剖式评测。结果导向的评估范式仅关注最终准确率。这种黑盒方法忽视了中间推理过程,极大弱化了评测的优化指导意义。
|
| 44 |
|
| 45 |
## 🏗️ 本仓库代码架构
|
| 46 |
|
|
@@ -270,7 +270,7 @@ docker ps -a --filter "ancestor=${IMAGE}" --format "{{.Names}}\t{{.Status}}"
|
|
| 270 |
|
| 271 |
## 📊 数据集统计信息
|
| 272 |
|
| 273 |
-
DataClawBench数据并非来源于合成样本或教学示例,而是基于发布团队在中国企业、产业与政策研究中的长期一线数据积累与行业洞察。当前版本以 2022 年相关数据为主,经过必要脱敏处理后构建任务,尽可能避免模型知识泄露且保留真实业务中的信息噪声与数据摩擦。任务撰写与标注由中山大学岭南学院专业团队完成,兼顾学术规范与业务可用性。
|
| 274 |
|
| 275 |
### 🗂️ 数据环境统计信息
|
| 276 |
|
|
@@ -336,7 +336,7 @@ DataClawBench数据并非来源于合成样本或教学示例,而是基于发
|
|
| 336 |
</tbody>
|
| 337 |
</table>
|
| 338 |
|
| 339 |
-
> 在任务执行层面,智能体通常需要在多文件间完成实体对齐、跨表关联、口径归一与聚合计算,而非单文件查值;必要时还需结合 `internal_metrics.csv` 中的业务约定。这也是
|
| 340 |
|
| 341 |
### 📋 任务统计信息
|
| 342 |
|
|
|
|
| 29 |
|
| 30 |
DataClawBench 正是面向这一变化而设计。它评测的不是脱离执行环境的抽象能力,而是在真实数据条件、明确任务约束和可复现执行协议下,OpenClaw 类端到端智能体完成数据分析任务的实际表现。
|
| 31 |
|
| 32 |
+
## 🔍 DataClawBench 是什么
|
| 33 |
|
| 34 |
DataClawBench 是一个面向真实复杂数据环境的过程导向数据分析任务评测基准。核心目标并非仅仅是衡量智能体在数据分析任务上的最终表现,而是旨在将其作为一个现实高保真的测试平台,同时细粒度地评估智能体在面临真实复杂环境和多步推理时的演化过程。
|
| 35 |
|
| 36 |
+
DataClawBench 大规模地模拟了高噪声、弱语义、跨领域的真实世界数据环境。通过金融领域与计算机领域的人类专家撰写复杂数据分析任务问题,并由人类专家和AI辅助交叉核验提供每个任务的过程性标注和唯一客观答案。其中,过程性标注包括任务里程碑、人工订正参考轨迹和证据数据来源。DataClawBench 采用 OpenClaw 作为统一 agent 框架。
|
| 37 |
|
| 38 |
|
| 39 |
## 🎯 为什么选择 DataClawBench
|
| 40 |
|
| 41 |
- **从理想化的假设数据环境到非理想化的真实数据环境。** DataClawBench 包含混合结构和非结构化数据,覆盖企业画像、企业经营状态、区域产业统计、全国行业统计与政策文本等资源。所有信息均来自真实世界采集,并面临指标缺失、口径不一致、命名不一致等摩擦;任务面对的是现实风格的数据环境,而非过度清洗后的单表查值。
|
| 42 |
+
- **从单点静态查询到多步动态推理。** DataClawBench 任务通常要求智能体完成多阶段操作链,而不是一次性命中答案。对 agent 的挑战不仅来自检索,更来自跨源整合、指标构造、聚合计算与格式约束执行。
|
| 43 |
+
- **从结果导向评估到过程导向评估。** DataClawBench 超越单纯的结果准确率评估,对智能体的运行演化过程进行解剖式评测。结果导向的评估范式仅关注最终准确率。这种黑盒方法忽视了中间推理过程,极大弱化了评测的优化指导意义。
|
| 44 |
|
| 45 |
## 🏗️ 本仓库代码架构
|
| 46 |
|
|
|
|
| 270 |
|
| 271 |
## 📊 数据集统计信息
|
| 272 |
|
| 273 |
+
DataClawBench 数据并非来源于合成样本或教学示例,而是基于发布团队在中国企业、产业与政策研究中的长期一线数据积累与行业洞察。当前版本以 2022 年相关数据为主,经过必要脱敏处理后构建任务,尽可能避免模型知识泄露且保留真实业务中的信息噪声与数据摩擦。任务撰写与标注由中山大学岭南学院专业团队完成,兼顾学术规范与业务可用性。
|
| 274 |
|
| 275 |
### 🗂️ 数据环境统计信息
|
| 276 |
|
|
|
|
| 336 |
</tbody>
|
| 337 |
</table>
|
| 338 |
|
| 339 |
+
> 在任务执行层面,智能体通常需要在多文件间完成实体对齐、跨表关联、口径归一与聚合计算,而非单文件查值;必要时还需结合 `internal_metrics.csv` 中的业务约定。这也是 DataClawBench 用于评估真实场景数据理解与推理能力的核心价值。
|
| 340 |
|
| 341 |
### 📋 任务统计信息
|
| 342 |
|