Release v1.0: add dataset card and unify MIT license for Hugging Face

.dockerignore

@@ -0,0 +1,16 @@
+.git
+.github
+.cursor
+.venv
+__pycache__
+*.pyc
+*.pyo
+*.pyd
+*.log
+results
+report
+tmp
+output
+Images
+agent-tools
+*.plan.md

.env.example

@@ -0,0 +1,53 @@
+# === Docker ===
+DOCKER_IMAGE=dataclaw:0.1.0
+GATEWAY_PORT=3333
+TMP_WORKSPACE=/tmp_workspace
+
+# === Benchmark ===
+DEFAULT_MODEL=
+JUDGE_MODEL=openrouter/anthropic/claude-opus-4.5
+DEFAULT_PARALLEL=1
+OUTPUT_SUBDIR=output
+BENCHMARK_TIMEOUT_MULTIPLIER=1.0
+BENCHMARK_RUNS=1
+
+# === Web search (OpenClaw web_search tool) ===
+# Register at https://brave.com/search/api/ (Search plan), then set the key.
+# OpenClaw docs: https://docs.openclaw.ai/tools/brave-search
+BRAVE_API_KEY=
+# Optional tuning (defaults match OpenClaw docs)
+# BRAVE_WEB_SEARCH_MAX_RESULTS=5
+# BRAVE_WEB_SEARCH_TIMEOUT_SECONDS=30
+
+# method 1: use openrouter
+# === Auth: OpenRouter ===
+OPENROUTER_API_KEY=
+
+# method 2: use custom api
+# === Auth: Custom OpenAI-compatible API (optional) ===
+# OPENCLAW_CUSTOM_BASE_URL=https://your-api-url/v1
+# OPENCLAW_CUSTOM_API_KEY=
+# OPENCLAW_CUSTOM_MODEL_ID=
+
+# === Auth: Judge Custom API (optional, separate endpoint for judge model) ===
+# JUDGE_CUSTOM_BASE_URL=https://your-judge-api-url/v1
+# JUDGE_CUSTOM_API_KEY=
+# JUDGE_CUSTOM_MODEL_ID=
+
+# === Agent model capabilities (cost unit: USD per 1M tokens) ===
+# OPENCLAW_MODEL_CONTEXT_WINDOW=128000
+# OPENCLAW_MODEL_MAX_TOKENS=16384
+# OPENCLAW_MODEL_COST_INPUT=0
+# OPENCLAW_MODEL_COST_OUTPUT=0
+# OPENCLAW_MODEL_COST_CACHE_READ=0
+# OPENCLAW_MODEL_COST_CACHE_WRITE=0
+
+# === Judge model capabilities (cost unit: USD per 1M tokens) ===
+# JUDGE_MODEL_CONTEXT_WINDOW=128000
+# JUDGE_MODEL_MAX_TOKENS=16384
+# JUDGE_MODEL_COST_INPUT=0
+# JUDGE_MODEL_COST_OUTPUT=0
+# JUDGE_MODEL_COST_CACHE_READ=0
+# JUDGE_MODEL_COST_CACHE_WRITE=0
+
+

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+assets/database/**/*.csv filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,26 @@
+FROM ubuntu:22.04
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    OPENCLAW_VERSION=2026.3.24
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ca-certificates \
+    curl \
+    git \
+    python3 \
+    python3-pip \
+    bash \
+    tini \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN curl -fsSL https://deb.nodesource.com/setup_22.x | bash - \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends nodejs \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN npm install -g "openclaw@${OPENCLAW_VERSION}"
+
+RUN mkdir -p /tmp_workspace /root/.openclaw/workspace
+
+ENTRYPOINT ["tini", "--"]
+CMD ["tail", "-f", "/dev/null"]

LICENSE

@@ -0,0 +1,22 @@
+MIT License
+
+Copyright (c) 2026 The DataClaw Authors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+

README.md

@@ -1,3 +1,373 @@
 ---
 license: mit
+task: data-analysis
+task_categories:
+  - question-answering
+  - tabular-question-answering
 ---
+
+<div align="center">
+
+<h1>DataClaw</h1>
+
+<img src="logo.png" alt="DataClaw Logo" width="220"/>
+
+<br/>
+<br/>
+
+[![🏆 Leaderboard](https://img.shields.io/badge/🏆_Leaderboard-DataClaw-red)](https://gtmllab.github.io/DataClaw/)
+![Tasks](https://img.shields.io/badge/Tasks-492-blue)
+![Categories](https://img.shields.io/badge/Categories-7-green)
+
+</div>
+
+> A data-analysis benchmark for OpenClaw-style end-to-end agents. Every task is grounded in real-world data and has a single objective gold answer.
+
+
+
+[简体中文](README.zh-CN.md)
+
+## 🌊 Data Analysis Tasks Are Changing in the OpenClaw Era
+
+With the emergence of end-to-end agents like OpenClaw, data analysis is no longer equivalent to static QA — "read a passage, output one answer." Real-world data analysis tasks often require agents to locate evidence across heterogeneous files, filter and join entities across tables, perform statistical and normalization calculations, verify intermediate results, and strictly follow output constraints.
+
+This means the core difficulty of a benchmark has shifted from answer generation alone to full agent-driven execution. A truly valuable data-analysis benchmark must test not only whether the final answer is correct, but also whether the agent can reliably complete a series of steps — retrieval, filtering, computation, verification, and constraint compliance — in complex data environments.
+
+DataClaw is designed for exactly this shift. It evaluates not abstract capability divorced from execution, but how OpenClaw-style end-to-end agents actually perform on data analysis tasks under real data conditions, explicit task constraints, and a reproducible execution protocol.
+
+## 🔍 What Is DataClaw?
+
+DataClaw 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.
+
+DataClaw 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. DataClaw adopts OpenClaw as its unified agent framework.
+
+
+## 🎯 Why DataClaw?
+
+- **From idealized data environments to imperfect real-world data environments.** DataClaw contains a mix of structured and unstructured data, covering enterprise profiles, business operating status, regional industry statistics, national industry statistics, and policy texts. All data is collected from the real world and comes with friction such as missing indicators, inconsistent definitions, and inconsistent naming. Tasks face realistic data environments, not over-cleaned single-table lookups.
+- **From single-shot static queries to multi-step dynamic reasoning.** DataClaw tasks typically require agents to complete a multi-stage chain of operations rather than producing a one-shot answer. The challenge for agents comes not only from retrieval but also from cross-source integration, metric construction, aggregation computation, and format constraint compliance.
+- **From outcome-oriented evaluation to process-oriented evaluation.** DataClaw goes beyond simple outcome-accuracy evaluation and dissects how the agent's execution unfolds at fine granularity. Outcome-oriented evaluation paradigms focus only on final accuracy. This black-box approach ignores intermediate reasoning and provides little actionable signal for guiding optimization.
+
+## 🏗️ Repository Layout
+
+
+Key directories and scripts:
+
+- `assets/database/`: Benchmark data files, injected wholesale into the container workspace at run time. The root contains `internal_metrics.csv` (internal business-logic knowledge base); `enterprise/`, `industry/`, and `policy/` hold the three theme-domain datasets.
+- `assets/qa_raw/`: Raw task source files.
+- `assets/qa_gold/`: Minimized gold files derived from `qa_raw`.
+- `tasks/`: Generated OpenClaw task spec files.
+- `dataclaw/build_tasks.py`: Builder that produces `qa_gold` and `tasks/` from `qa_raw`.
+- `dataclaw/eval/run_batch.py`: Host-side evaluation orchestrator; one isolated container per task.
+- `dataclaw/utils/docker_utils.py`: Container lifecycle management, OpenClaw onboarding, and model configuration.
+- `dataclaw/utils/grading.py`: Outcome scoring (LLM-judged Acc).
+- `dataclaw/utils/process_grading.py`: Process scoring (EE on correct tasks; GPR / TPE on incorrect tasks).
+- `script/docker_save_image.sh`: Image build and export script.
+
+### ⚙️ Evaluation Lifecycle
+
+Each evaluation task runs in its own Docker container. The host orchestrator manages the full lifecycle:
+
+```text
+Host (dataclaw/eval/run_batch.py)
+  |
+  +-- For each task (parallel via --parallel N):
+      1. docker run   -> start isolated container
+      2. docker cp    -> inject workspace files
+      3. docker exec  -> OpenClaw onboard
+      4. docker exec  -> start gateway (background)
+      5. docker exec  -> set model and run agent
+      6. docker exec  -> run llm_judge scoring
+      7. docker cp    -> collect logs and results
+      8. docker rm    -> remove container
+```
+
+
+## 🚀 User Quick Start
+
+### 1. Obtain the Pre-built Image
+
+Download the pre-built image archive from Releases and load it:
+
+```bash
+docker load -i <dataclaw-image-archive>.tar
+```
+
+After loading, confirm the local image tag matches `DOCKER_IMAGE` in `.env`.
+
+### 2. Clone This Repository
+
+```bash
+git clone <repository-url>
+cd <repository-dir>
+```
+
+Use the actual repository URL shown on the GitHub page.
+
+### 3. Install Python Dependencies
+
+```bash
+pip install pyyaml python-dotenv
+```
+
+> `pyproject.toml` requires Python `>=3.10`. For a fuller local dev setup, install additional dev dependencies as you prefer.
+
+### 4. Configure Environment
+
+Copy the template:
+
+```bash
+cp .env.example .env
+```
+
+Edit `.env` and pay attention to at least the following fields:
+
+| Variable | Required | Description |
+| --- | --- | --- |
+| `DEFAULT_MODEL` | Yes | Model under test, e.g. `openrouter/anthropic/claude-sonnet-4.6` |
+| `OPENROUTER_API_KEY` | One of two | Used when the main model or judge is called via OpenRouter |
+| `OPENCLAW_CUSTOM_BASE_URL` + `OPENCLAW_CUSTOM_API_KEY` | One of two | Custom OpenAI-compatible API |
+| `OPENCLAW_CUSTOM_MODEL_ID` | No | Explicit model id at the custom provider for the main model |
+| `JUDGE_MODEL` | No | Judge model; default in `.env.example` |
+| `JUDGE_CUSTOM_BASE_URL` + `JUDGE_CUSTOM_API_KEY` | No | Separate custom endpoint for the judge |
+| `JUDGE_CUSTOM_MODEL_ID` | No | Explicit model id for the judge custom endpoint |
+| `DOCKER_IMAGE` | No | Local image tag; must match the loaded image |
+
+#### 🔌 Custom OpenAI-compatible API
+
+If you do not use OpenRouter, set in `.env`:
+
+```bash
+OPENCLAW_CUSTOM_BASE_URL=https://your-api-url/v1
+OPENCLAW_CUSTOM_API_KEY=your_api_key
+OPENCLAW_CUSTOM_MODEL_ID=your-provider/your-model
+DEFAULT_MODEL=your-provider/your-model
+```
+
+If the API runs on the host:
+
+```bash
+OPENCLAW_CUSTOM_BASE_URL=http://host.docker.internal:8000/v1
+```
+
+When the judge uses a separate endpoint:
+
+```bash
+JUDGE_CUSTOM_BASE_URL=https://your-judge-api-url/v1
+JUDGE_CUSTOM_API_KEY=your_judge_api_key
+JUDGE_CUSTOM_MODEL_ID=your-provider/your-judge-model
+```
+
+##### Common Auth Setups (Main Model vs Judge)
+
+| Scenario | Main model | Judge | Required config |
+| --- | --- | --- | --- |
+| A | Custom API | OpenRouter | `OPENCLAW_CUSTOM_*` + `OPENROUTER_API_KEY` |
+| B | OpenRouter | OpenRouter | `OPENROUTER_API_KEY` |
+| C | Custom API | Custom API (separate endpoint) | `OPENCLAW_CUSTOM_*` + `JUDGE_CUSTOM_*` |
+
+### 5. Run Evaluation
+
+```bash
+# Run all tasks
+python dataclaw/eval/run_batch.py --model openrouter/anthropic/claude-sonnet-4.6
+
+# Run selected tasks
+python dataclaw/eval/run_batch.py --model ... --suite task_001,task_002
+
+# Run in parallel
+python dataclaw/eval/run_batch.py --model ... --parallel 4
+
+# Run a single task file
+python dataclaw/eval/run_batch.py --task tasks/task_001_xxx.md
+
+# Or use the convenience script (reads DEFAULT_MODEL from .env)
+bash script/run.sh
+```
+
+### CLI Options
+
+| Flag | Default | Description |
+| --- | --- | --- |
+| `--model` / `-m` | `DEFAULT_MODEL` in `.env` | Model under test |
+| `--judge` | `JUDGE_MODEL` in `.env` | Judge model |
+| `--suite` / `-s` | `all` | `"all"` or comma-separated task IDs |
+| `--task` / `-t` | — | Path to a single `task.md` |
+| `--parallel` / `-p` | `1` | Parallel container count |
+| `--timeout-multiplier` | `1.0` | Scale all task timeouts |
+| `--runs` | `1` | Repeat runs per task |
+| `--resume` | — | Resume from last interrupted run |
+| `--verbose` / `-v` | — | Enable verbose logging |
+
+### 6. Results
+
+After a run completes, results are saved under `output/<task_id>/<model_timestamp_runid>/`:
+
+```text
+output/<task_id>/<suffix>/
+├── score.json               # outcome score (Acc)
+├── process_score.json       # process scores (EE / GPR / TPE)
+├── usage.json               # token usage, cost, elapsed time
+├── agent.log                # agent execution log
+├── gateway.log              # gateway log
+├── chat.jsonl               # full conversation record
+├── judge_chat.jsonl         # outcome-judge conversation
+└── judge_process_chat.jsonl # process-judge conversation
+```
+
+A global summary is written to:
+
+```text
+output/summary_<model>.json
+```
+
+### 7. Grading Rules
+
+DataClaw scores each run along **four metrics**.
+
+| Metric | Definition | Scope | Direction |
+| --- | --- | --- | --- |
+| **Acc** | LLM-judge semantic match between predicted answer â and gold answer a; multi-subquestion tasks return the normalized mean over the L sub-answers. | All tasks | ↑ |
+| **EE** | Execution Efficiency = N / T, where N is the gold reference step count and T is the agent's actual step count. EE > 1 means the agent solved it in fewer steps than the gold trajectory. | Correct tasks only | ↑ |
+| **GPR** | Goal Progress Rate = (1/M) Σⱼ 𝕀(mⱼ achieved); fraction of M annotated milestones the agent reaches anywhere in its trajectory. Captures partial process credit when the final answer is wrong. | Incorrect tasks only | ↑ |
+| **TPE** | Temporal Progress Efficiency = (Σⱼ 𝕀(mⱼ) · γ^max(tⱼ − N, 0)) / Σⱼ 𝕀(mⱼ), γ = 0.9; averages an exponential temporal-decay factor over the milestones the agent did achieve. Milestones reached by step N contribute 1; later ones decay. TPE = 1 means all achieved milestones were on time, lower values indicate later concentration. Range [0, 1]. | Incorrect tasks only | ↑ |
+
+### 8. Resume After Interruption
+
+Long batch evaluations may be interrupted unexpectedly. The evaluation framework automatically saves a progress file after each task completes:
+
+```text
+output/progress_<model>.json
+```
+
+Simply append `--resume` to the original command to resume:
+
+```bash
+# Original run (interrupted midway)
+python dataclaw/eval/run_batch.py --model openrouter/anthropic/claude-sonnet-4.6 --suite all
+
+# Resume (keep other arguments the same)
+python dataclaw/eval/run_batch.py --model openrouter/anthropic/claude-sonnet-4.6 --suite all --resume
+```
+
+On resume, the framework automatically verifies that `--model`, `--suite`, and `--runs` match the previous run; if they don't, it exits with an error. Once all tasks are completed, the progress file is removed automatically.
+
+To discard previous progress and start fresh, simply delete the progress file:
+
+```bash
+rm output/progress_<model>.json
+```
+
+### 9. Cleanup
+
+Interrupted runs may leave behind uncleaned containers. Clean them up as follows:
+
+```bash
+IMAGE=<your-docker-image-tag>
+docker ps -a --filter "ancestor=${IMAGE}" -q | xargs -r docker rm -f
+```
+
+Preview containers that would be removed:
+
+```bash
+IMAGE=<your-docker-image-tag>
+docker ps -a --filter "ancestor=${IMAGE}" --format "{{.Names}}\t{{.Status}}"
+```
+
+## 📊 Dataset Statistics
+
+DataClaw's data does not come from synthetic samples or teaching examples; it is built on the publishing team's long-term, front-line data accumulation and industry insights from research on Chinese enterprises, industries, and policies. The current version is mainly based on data from 2022. After necessary de-identification, tasks are constructed to avoid model knowledge leakage as much as possible while preserving the information noise and data friction found in real business settings. Task authoring and annotation are conducted by a professional team from Lingnan College, Sun Yat-sen University, balancing academic rigor and practical usability.
+
+### 🗂️ Data Environment Statistics
+
+Under a theme-domain view and business-oriented taxonomy, the current data environment is organized into **3 theme domains**: **Enterprise**, **Industry**, and **Policy**; subcategories cover enterprise profiles and regional profiles, enterprise core competitiveness, business operating status, regional/national industry statistics, policy releases, and full policy texts — closely aligned with real research and consulting workflows. The 3 theme domains contain **17 independent data sources** (each data file counts as one source, all mounted under `assets/database/`): **17** are placed under theme subdirectories `enterprise/`, `industry/`, `policy/`; in addition, **1** root-level file, `internal_metrics.csv`, serves as an **internal business-logic knowledge base** and does not belong to any theme domain. Details below.
+
+| Dimension | Value | Notes |
+| --- | --- | --- |
+| Theme domains | 3 | Enterprise, Industry, Policy |
+| Secondary themes | 7 | Enterprise ×3 (profiles, core competitiveness, business status), Industry ×2 (regional industry, national industry), Policy ×2 (release status, full text) |
+| Total data sources | 17 | Injected into the container workspace at run time |
+| Format | CSV | Primarily CSV; includes both structured fields and unstructured long-text content |
+| Time span | Mainly concentrated in 2022 | Statistical periods vary across sources |
+
+**The 17 data sources by theme domain and secondary theme**
+
+<table>
+<thead>
+<tr><th>Theme domain</th><th>Secondary theme</th><th align="right">Sources</th><th>Files</th></tr>
+</thead>
+<tbody>
+<tr>
+<td style="white-space:nowrap">Enterprise</td>
+<td style="white-space:nowrap">Enterprise profiles</td>
+<td align="right">5</td>
+<td><code>enterprise/company_profile.csv</code><br><code>enterprise/company_profile_as.csv</code><br><code>enterprise/company_profile_eu.csv</code><br><code>enterprise/company_profile_na.csv</code><br><code>enterprise/company_profile_oc.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">Enterprise</td>
+<td style="white-space:nowrap">Core competitiveness</td>
+<td align="right">1</td>
+<td><code>enterprise/company_core.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">Enterprise</td>
+<td style="white-space:nowrap">Business status</td>
+<td align="right">3</td>
+<td><code>enterprise/company_operation_status.csv</code><br><code>enterprise/company_operation_status_detail.csv</code><br><code>enterprise/company_operation_yearly_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">Industry</td>
+<td style="white-space:nowrap">Regional industry</td>
+<td align="right">3</td>
+<td><code>industry/regional_industry_status.csv</code><br><code>industry/regional_industry_status_detail.csv</code><br><code>industry/regional_industry_yearly_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">Industry</td>
+<td style="white-space:nowrap">National industry</td>
+<td align="right">3</td>
+<td><code>industry/national_industry_status.csv</code><br><code>industry/national_industry_status_detail.csv</code><br><code>industry/national_industry_yearly_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">Policy</td>
+<td style="white-space:nowrap">Policy release status</td>
+<td align="right">1</td>
+<td><code>policy/policy_release_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">Policy</td>
+<td style="white-space:nowrap">Full policy text</td>
+<td align="right">1</td>
+<td><code>policy/policy_resource.csv</code></td>
+</tr>
+</tbody>
+</table>
+
+> 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 DataClaw for evaluating real-scenario data understanding and reasoning capabilities.
+
+### 📋 Task Statistics
+
+The current version contains **492** tasks across **7** categories, with an overall difficulty distribution of **131 easy / 286 medium / 75 hard**.
+
+| Category code | Meaning | Count | Difficulty split |
+| --- | --- | --- | --- |
+| `enterprise_industry_analysis` | Enterprise–industry analysis | 226 | easy 115 / medium 111 |
+| `enterprise_industry_policy_analysis` | Enterprise–industry–policy linkage analysis | 76 | easy 10 / medium 66 |
+| `comprehensive_decision` | Comprehensive decision | 70 | easy 6 / medium 45 / hard 19 |
+| `international_comparison` | International comparison | 39 | medium 25 / hard 14 |
+| `hypothesis_verification` | Hypothesis verification | 29 | medium 14 / hard 15 |
+| `industry_planning` | Industry planning | 28 | medium 14 / hard 14 |
+| `risk_assessment` | Risk assessment | 24 | medium 11 / hard 13 |
+
+> Except for the 39 `international_comparison` tasks, all others are explicitly restricted in the current task spec to use only `./database/`, with no web search.
+
+## 🙏 Acknowledgements
+
+DataClaw is jointly released by Prof. Chuan Chen's team at the School of Computer Science, Sun Yat-sen University, and the Southern Weekly Sci-Tech Power Research Center. We sincerely thank the Southern Weekly Sci-Tech Power Research Center for providing invaluable data and tremendous support.
+
+This project also builds on excellent open-source agent ecosystems. We gratefully acknowledge:
+
+ [WildClawBench](https://github.com/InternLM/WildClawBench)
+
+ [Claw-Eval](https://github.com/claw-eval/claw-eval)
+
+ [PinchBench](https://github.com/pinchbench/skill)

README.zh-CN.md

@@ -0,0 +1,368 @@
+<div align="center">
+
+<h1>DataClaw</h1>
+
+<img src="logo.png" alt="DataClaw Logo" width="220"/>
+
+<br/>
+<br/>
+
+[![🏆 Leaderboard](https://img.shields.io/badge/🏆_Leaderboard-DataClaw-red)](https://gtmllab.github.io/DataClaw/)
+[![🤗 HuggingFace](https://img.shields.io/badge/🤗_HuggingFace-Dataset-yellow)](https://huggingface.co/datasets/GTMLLab/DataClaw)
+![Tasks](https://img.shields.io/badge/Tasks-492-blue)
+![Categories](https://img.shields.io/badge/Categories-7-green)
+
+</div>
+
+> 面向 OpenClaw 类端到端智能体的数据分析评测集，所有任务来自真实数据环境，且具备唯一客观答案。
+
+
+
+[English README](README.md)
+
+## 🌊 OpenClaw 时代的数据分析任务正在变化
+
+随着 OpenClaw 一类端到端智能体的出现，数据分析任务已经不再等价于“读一段文本，再输出一个答案”的静态问答问题。真实世界中的数据分析任务往往要求智能体在异构文件中定位证据、跨表筛选与关联实体、执行统计与归一化计算、校验中间结果，并严格遵循输出约束。
+
+这意味着评测基准的核心难点，已经从单一的答案生成，转向完整的智能体驱动执行。一个真正有价值的数据分析评测基准，不仅要考察最终答案是否正确，还要考察智能体是否能够在复杂数据环境中稳定完成检索、筛选、计算、验证与约束遵循等一系列步骤。
+
+DataClaw 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。
+
+## 🔍 DataClaw 是什么
+
+DataClaw 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。
+
+DataClaw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。DataClaw 采用 OpenClaw 作为统一 agent 框架。
+
+
+## 🎯 为什么选择 DataClaw
+
+- **从理想化的假设数据环境到非理想化的真实数据环境。** DataClaw 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。
+- **从单点静态查询到多步动态推理。** DataClaw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。
+- **从结果导向评估到过程导向评估。** DataClaw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。
+
+## 🏗️ 本仓库代码架构
+
+
+关键目录与脚本职责如下：
+
+- `assets/database/`：评测数据文件，运行时会整体注入到容器工作区。根目录含 `internal_metrics.csv`（业务逻辑内部知识库）；`enterprise/`、`industry/`、`policy/` 为三大主题域数据。
+- `assets/qa_raw/`：原始任务源文件。
+- `assets/qa_gold/`：由 `qa_raw` 归约得到的精简 gold 文件。
+- `tasks/`：生成后的 OpenClaw task 规范文件。
+- `dataclaw/build_tasks.py`：从 `qa_raw` 生成 `qa_gold` 与 `tasks/` 的构建器。
+- `dataclaw/eval/run_batch.py`：宿主机侧评测编排器，负责每任务独立容器执行。
+- `dataclaw/utils/docker_utils.py`：容器生命周期管理、OpenClaw onboarding 与模型配置。
+- `dataclaw/utils/grading.py`：结果评分（LLM-judge 计算 Acc）。
+- `dataclaw/utils/process_grading.py`：过程评分（correct 任务计 EE；incorrect 任务计 GPR / TPE）。
+- `script/docker_save_image.sh`：镜像构建与导出脚本。
+
+### ⚙️ 评测执行生命周期
+
+每个评测任务都在独立 Docker 容器中运行。宿主机编排器管理完整生命周期：
+
+```text
+宿主机 (dataclaw/eval/run_batch.py)
+  |
+  +-- 对每个任务（通过 --parallel N 并行）：
+      1. docker run   -> 启动隔离容器
+      2. docker cp    -> 注入工作区文件
+      3. docker exec  -> OpenClaw onboard
+      4. docker exec  -> 启动 gateway（后台）
+      5. docker exec  -> 设置模型并运行 agent
+      6. docker exec  -> 调用 llm_judge 评分
+      7. docker cp    -> 收集日志与结果
+      8. docker rm    -> 清理容器
+```
+
+
+## 🚀 使用者快速开始
+
+### 1. 获取预构建镜像
+
+从 Releases 下载预构建镜像压缩包并加载：
+
+```bash
+docker load -i <dataclaw-image-archive>.tar
+```
+
+加载后，请确认本地镜像 tag 与 `.env` 中的 `DOCKER_IMAGE` 一致。
+
+### 2. 克隆本仓库
+
+```bash
+git clone <repository-url>
+cd <repository-dir>
+```
+
+实际仓库地址请以当前 GitHub 页面为准。
+
+### 3. 安装 Python 依赖
+
+```bash
+pip install pyyaml python-dotenv
+```
+
+> `pyproject.toml` 要求 Python `>=3.10`。如需更完整的本地开发环境，也可以按你的习惯额外安装开发依赖。
+
+### 4. 配置环境
+
+复制模板：
+
+```bash
+cp .env.example .env
+```
+
+编辑 `.env`，至少关注以下字段：
+
+| 变量 | 必填 | 说明 |
+| --- | --- | --- |
+| `DEFAULT_MODEL` | 是 | 待评测主模型，例如 `openrouter/anthropic/claude-sonnet-4.6` |
+| `OPENROUTER_API_KEY` | 二选一 | 当主模型或 Judge 通过 OpenRouter 调用时使用 |
+| `OPENCLAW_CUSTOM_BASE_URL` + `OPENCLAW_CUSTOM_API_KEY` | 二选一 | 自定义 OpenAI-compatible API |
+| `OPENCLAW_CUSTOM_MODEL_ID` | 否 | 自定义主模型在 provider 中的显式 model id |
+| `JUDGE_MODEL` | 否 | Judge 模型，默认值见 `.env.example` |
+| `JUDGE_CUSTOM_BASE_URL` + `JUDGE_CUSTOM_API_KEY` | 否 | Judge 使用独立自定义 endpoint 时填写 |
+| `JUDGE_CUSTOM_MODEL_ID` | 否 | Judge 自定义 endpoint 的显式 model id |
+| `DOCKER_IMAGE` | 否 | 本地镜像 tag，需与实际加载的镜像一致 |
+
+#### 🔌 自定义 OpenAI-compatible API
+
+如果不使用 OpenRouter，可以在 `.env` 中设置：
+
+```bash
+OPENCLAW_CUSTOM_BASE_URL=https://your-api-url/v1
+OPENCLAW_CUSTOM_API_KEY=your_api_key
+OPENCLAW_CUSTOM_MODEL_ID=your-provider/your-model
+DEFAULT_MODEL=your-provider/your-model
+```
+
+如果 API 运行在宿主机上：
+
+```bash
+OPENCLAW_CUSTOM_BASE_URL=http://host.docker.internal:8000/v1
+```
+
+Judge 使用独立 endpoint 时：
+
+```bash
+JUDGE_CUSTOM_BASE_URL=https://your-judge-api-url/v1
+JUDGE_CUSTOM_API_KEY=your_judge_api_key
+JUDGE_CUSTOM_MODEL_ID=your-provider/your-judge-model
+```
+
+##### 主模型与 Judge 的常见认证组合
+
+| 场景 | 主模型 | Judge | 所需配置 |
+| --- | --- | --- | --- |
+| A | 自定义 API | OpenRouter | `OPENCLAW_CUSTOM_*` + `OPENROUTER_API_KEY` |
+| B | OpenRouter | OpenRouter | `OPENROUTER_API_KEY` |
+| C | 自定义 API | 自定义 API（独立 endpoint） | `OPENCLAW_CUSTOM_*` + `JUDGE_CUSTOM_*` |
+
+### 5. 运行评测
+
+```bash
+# 运行全部任务
+python dataclaw/eval/run_batch.py --model openrouter/anthropic/claude-sonnet-4.6
+
+# 运行指定任务
+python dataclaw/eval/run_batch.py --model ... --suite task_001,task_002
+
+# 并行运行
+python dataclaw/eval/run_batch.py --model ... --parallel 4
+
+# 运行单个任务文件
+python dataclaw/eval/run_batch.py --task tasks/task_001_xxx.md
+
+# 或使用便捷脚本（从 .env 读取 DEFAULT_MODEL）
+bash script/run.sh
+```
+
+### CLI 选项
+
+| 参数 | 默认值 | 说明 |
+| --- | --- | --- |
+| `--model` / `-m` | `.env` 中的 `DEFAULT_MODEL` | 待评测模型 |
+| `--judge` | `.env` 中的 `JUDGE_MODEL` | Judge 模型 |
+| `--suite` / `-s` | `all` | `"all"` 或逗号分隔的任务 ID |
+| `--task` / `-t` | — | 单个 `task.md` 文件路径 |
+| `--parallel` / `-p` | `1` | 并行容器数 |
+| `--timeout-multiplier` | `1.0` | 缩放所有任务超时时间 |
+| `--runs` | `1` | 每个任务的重复运行次数 |
+| `--resume` | — | 从上次中断处恢复运行 |
+| `--verbose` / `-v` | — | 启用详细日志 |
+
+### 6. 结果
+
+运行完成后，结果保存在 `output/<task_id>/<model_timestamp_runid>/` 下：
+
+```text
+output/<task_id>/<suffix>/
+├── score.json               # 结果评分（Acc）
+├── process_score.json       # 过程评分（EE / GPR / TPE）
+├── usage.json               # token 统计、费用、耗时
+├── agent.log                # agent 执行日志
+├── gateway.log              # gateway 日志
+├── chat.jsonl               # 完整对话记录
+├── judge_chat.jsonl         # outcome-judge 对话
+└── judge_process_chat.jsonl # process-judge 对话
+```
+
+全局汇总会写入：
+
+```text
+output/summary_<model>.json
+```
+
+### 7. 评分规则
+
+DataClaw 对每次运行从 **四个指标** 打分。
+
+| 指标 | 含义 | 计算范围 | 方向 |
+| --- | --- | --- | --- |
+| **Acc** | LLM-judge 评判预测答案 â 与标准答案 a 的语义一致性，含 L 个子问题时取归一化平均 | 全部任务 | ↑ |
+| **EE** | 执行效率 = N / T（N=gold 参考步数，T=agent 实际步数）。EE>1 表示比 gold 更省步 | 仅 correct 任务 | ↑ |
+| **GPR** | 目标进展率 = (1/M) Σⱼ 𝕀(mⱼ 达成)，agent 在轨迹中达成的里程碑占比，用于在答错时捕捉过程得分 | 仅 incorrect 任务 | ↑ |
+| **TPE** | 时序进展效率 = (Σⱼ 𝕀(mⱼ) · γ^max(tⱼ−N, 0)) / Σⱼ 𝕀(mⱼ)，γ=0.9；对 agent 已达成的里程碑做时序衰减平均，N 步内达成贡献 1，之后按 γ 指数衰减。所有已达成里程碑均在第 N 步内达成时 TPE=1，越晚 TPE 越低。取值范围 [0, 1]。 | 仅 incorrect 任务 | ↑ |
+
+### 8. 中断恢复
+
+长时间的批量评测可能因意外而中断。评测框架会在每个任务完成后自动保存进度文件：
+
+```text
+output/progress_<model>.json
+```
+
+只需在原命令后加上 `--resume` 即可恢复：
+
+```bash
+# 原始运行（中途中断）
+python dataclaw/eval/run_batch.py --model openrouter/anthropic/claude-sonnet-4.6 --suite all
+
+# 恢复运行（其余参数保持一致）
+python dataclaw/eval/run_batch.py --model openrouter/anthropic/claude-sonnet-4.6 --suite all --resume
+```
+
+恢复时，框架会自动校验 `--model`、`--suite`、`--runs` 是否与上次运行一致；若不一致，会直接报错退出。全部任务完成后，进度文件会被自动删除。
+
+如需放弃之前的进度并重新开始，删除对应进度文件即可：
+
+```bash
+rm output/progress_<model>.json
+```
+
+### 9. 清理
+
+如果运行被中断，可能会留下尚未清理的容器。可以按如下方式清理：
+
+```bash
+IMAGE=<your-docker-image-tag>
+docker ps -a --filter "ancestor=${IMAGE}" -q | xargs -r docker rm -f
+```
+
+预览会被移除的容器：
+
+```bash
+IMAGE=<your-docker-image-tag>
+docker ps -a --filter "ancestor=${IMAGE}" --format "{{.Names}}\t{{.Status}}"
+```
+
+## 📊 数据集统计信息
+
+DataClaw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。
+
+### 🗂️ 数据环境统计信息
+
+当前版本数据环境在主题域内，按业务口径归纳为 **3 个主题域**：**企业**、**产业**、**政策**；细分类目覆盖企业画像与分区画像、企业核心竞争力、经营状态、区域/全国产业统计、政策发布与政策原文等，贴近真实研究与咨询分析链路。3个主题域包含**17 个独立数据源**（每个数据文件计为一个数据源，统一挂载到 `assets/database/`）：其中 **17** 个按主题域归入子目录 `enterprise/`、`industry/`、`policy/`；另有 **1** 个根目录文件 `internal_metrics.csv`，作为**业务逻辑内部知识库**，不归属任一主题域。详细说明见下。
+
+| 维度 | 统计值 | 说明 |
+| --- | --- | --- |
+| 主题域 | 3 | 企业、产业、政策 |
+| 二级主题 | 7 | 企业 3（画像、核心竞争力、经营状态），产业 2（区域产业、全国行业），政策 2（发布情况、原文） |
+| 数据源总数 | 17 | 运行时整体注入容器工作区 |
+| 文件格式 | CSV | 以 CSV 为主；同时包含结构化字段与非结构化长文本型内容 |
+| 时间跨度 | 主要集中于2022年 | 不同数据源统计周期不一致 |
+
+**各主题域及二级主题下的 17 个数据源**
+
+<table>
+<thead>
+<tr><th>主题域</th><th>二级主题</th><th align="right">数据源数量</th><th>数据文件</th></tr>
+</thead>
+<tbody>
+<tr>
+<td style="white-space:nowrap">企业</td>
+<td style="white-space:nowrap">企业画像</td>
+<td align="right">5</td>
+<td><code>enterprise/company_profile.csv</code><br><code>enterprise/company_profile_as.csv</code><br><code>enterprise/company_profile_eu.csv</code><br><code>enterprise/company_profile_na.csv</code><br><code>enterprise/company_profile_oc.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">企业</td>
+<td style="white-space:nowrap">企业核心竞争力</td>
+<td align="right">1</td>
+<td><code>enterprise/company_core.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">企业</td>
+<td style="white-space:nowrap">企业经营状态</td>
+<td align="right">3</td>
+<td><code>enterprise/company_operation_status.csv</code><br><code>enterprise/company_operation_status_detail.csv</code><br><code>enterprise/company_operation_yearly_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">产业</td>
+<td style="white-space:nowrap">区域产业</td>
+<td align="right">3</td>
+<td><code>industry/regional_industry_status.csv</code><br><code>industry/regional_industry_status_detail.csv</code><br><code>industry/regional_industry_yearly_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">产业</td>
+<td style="white-space:nowrap">全国行业</td>
+<td align="right">3</td>
+<td><code>industry/national_industry_status.csv</code><br><code>industry/national_industry_status_detail.csv</code><br><code>industry/national_industry_yearly_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">政策</td>
+<td style="white-space:nowrap">政策发布情况</td>
+<td align="right">1</td>
+<td><code>policy/policy_release_status.csv</code></td>
+</tr>
+<tr>
+<td style="white-space:nowrap">政策</td>
+<td style="white-space:nowrap">政策原文</td>
+<td align="right">1</td>
+<td><code>policy/policy_resource.csv</code></td>
+</tr>
+</tbody>
+</table>
+
+> 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 DataClaw 用于评估真实场景数据理解与推理能力的核心价值。
+
+### 📋 任务统计信息
+
+当前版本共 492 个任务，覆盖 7 个类别；整体难度分布为 131 easy / 286 medium / 75 hard。
+
+| 类别代码 | 含义 | 任务数 | 难度分布 |
+| --- | --- | --- | --- |
+| `enterprise_industry_analysis` | 企业-行业分析 | 226 | easy 115 / medium 111 |
+| `enterprise_industry_policy_analysis` | 企业-产业-政策联动分析 | 76 | easy 10 / medium 66 |
+| `comprehensive_decision` | 综合决策 | 70 | easy 6 / medium 45 / hard 19 |
+| `international_comparison` | 国际比较 | 39 | medium 25 / hard 14 |
+| `hypothesis_verification` | 假设验证 | 29 | medium 14 / hard 15 |
+| `industry_planning` | 产业规划 | 28 | medium 14 / hard 14 |
+| `risk_assessment` | 风险评估 | 24 | medium 11 / hard 13 |
+
+> 除 `international_comparison` 的 39 个任务外，其余任务在当前 task 规范中都显式限制为仅使用 `./database/`，不依赖 web search。
+
+## 🙏 致谢
+
+DataClaw 由中山大学计算机学院陈川团队与南方周末科创力研究中心联合发布，真诚感谢南方周末科创力研究中心提供的宝贵数据和巨大支持。
+
+同时，本项目构建在优秀的开源智能体生态系统之上。我们衷心感谢以下项目：
+
+ [WildClawBench](https://github.com/InternLM/WildClawBench)
+
+ [Claw-Eval](https://github.com/claw-eval/claw-eval)
+
+ [PinchBench](https://github.com/pinchbench/skill)
+
+

assets/database/bilingual_translation_english_chinese.json

@@ -0,0 +1,340 @@
+{
+  "company_items": {
+    "海山昌工设备公司": [
+      "Haishan Chang Industrial Equipment Company",
+      "Haishan Changgong Equipment Company"
+    ],
+    "驰宏冶冶有色金属公司": "Chihong Yeye Non-ferrous Metals Co., Ltd",
+    "众白锦贸连锁公司": "Zhongbai Jinmao Chain Company",
+    "三三达腾重工公司": [
+      "Sansan Dateng Heavy Industry Company",
+      "Sansan Daten Heavy Industry Company"
+    ],
+    "华鲁润源科技股份公司": "Hualu Runyuan Technology Co., Ltd.",
+    "连机创机机床公司": "Lianji Chuangji Machine Tool Company",
+    "康盛安健生物医药公司": "Kangsheng Anjian Biopharmaceutical Company",
+    "恒逸昌化科技股份公司": "Hengyi Changhua Technology Co., Ltd.",
+    "东车科信系统公司": "Dongche Kexin Systems Company",
+    "众课软创软件公司": "Zhongke Ruanchuang Software Company",
+    "永惠昌达批发公司": "Yonghui Changda Wholesale Company",
+    "锌锗金泽材料公司": "Xin Ge Jinze Materials Company",
+    "健名安元医疗科技公司": "Jianming Anyuan Medical Technology Company",
+    "德西锦锦智能电气公司": "Dexi Jinjin Intelligent Electrical Company",
+    "永惠泽盛连锁公司": "Yonghui Zesheng Chain Company",
+    "步步盛锦商贸公司": "Bubusheng Jin Commerce Company",
+    "联花通泽商贸公司": [
+      "Lianhua Tongze Commerce Company",
+      "Lianhua Tongze Trading Company"
+    ],
+    "恒逸昌化精细化工公司": "Hengyi Changhua Fine Chemical Company",
+    "恒逸源锦精细化工公司": "Hengyi Yuanjin Fine Chemical Company",
+    "荣盛锦盛化学公司": "Rongsheng Jinsheng Chemical Company",
+    "物丽昌源批发公司": [
+      "Wulichangyuan Wholesale Company",
+      "Wu Li Chang Yuan Wholesale Co., Ltd.",
+      "Wu Li Chang Yuan Pi Fa Co., Ltd."
+    ],
+    "果投泽源新能源公司": "Guotouzeyuan New Energy Company",
+    "华谊昌泽科技股份公司": "Huayichangze Technology Co., Ltd.",
+    "药石生康医疗器械公司": "Yaoshi Shenkang Medical Equipment Company",
+    "宝新软联软件公司": "Baoxin Ruanlian Software Company",
+    "快克创锦设备公司": "Kuaike Chuangjin Equipment Company",
+    "浪集软创信息技术公司": "Langji Ruanchuang Information Technology Company",
+    "以山生辰医疗科技公司": "Yishan Shengchen Medical Technology Company",
+    "华仁泰泽药业股份公司": "Huaren Taize Pharmaceutical Co., Ltd.",
+    "复河辰泽生物医药公司": "Fuhe Chenze Biopharmaceutical Company",
+    "临公航腾重工公司": "Lingong Hangteng Heavy Industry Company",
+    "恒逸盛盛科技股份公司": "Hengyi Shengsheng Technology Co., Ltd.",
+    "以山泽元药业股份公司": "Yishan Zeyuan Pharmaceutical Co., Ltd.",
+    "瑞行泰元医疗器械公司": "Ruiying Taiyuan Medical Equipment Co., Ltd.",
+    "青青锦饮食品公司": "Qingqing Jinyin Food Company",
+    "安步尚昌品牌公司": "Anbu Shangchang Brand Company",
+    "普各瑞健生物医药公司": "Puge Ruijian Biopharmaceutical Company",
+    "贝壳": "KE Holdings",
+    "万国数据": "GDS Holdings",
+    "北汽路远新能源汽车公司": "Bei Qi Lu Yuan Xin Neng Yuan Qi Che Co., Ltd.",
+    "康盛康健药业股份公司": "Kangsheng Kangjian Pharmaceutical Co., Ltd.",
+    "众集达昌铜业公司": "Zhong Ji Da Chang Tong Ye Co., Ltd.",
+    "众海工筑锦建筑设计公司": "Zhong Hai Gong Zhu Jin Jian Zhu She Ji Co., Ltd.",
+    "陵有色冶达资源公司": "Ling You Se Ye Da Zi Yuan Co., Ltd.",
+    "用丰信创科技公司": "Yong Feng Xin Chuang Ke Ji Co., Ltd.",
+    "润会数智系统公司": "Run Hui Shu Zhi Xi Tong Co., Ltd.",
+    "众通捷通运输公司": "Zhong Tong Jie Tong Yun Shu Co., Ltd.",
+    "云达航昌快递公司": "Yun Da Hang Chang Kuai Di Co., Ltd.",
+    "环星锦雅时尚公司": "Huan Xing Jin Ya Shi Shang Co., Ltd.",
+    "李丁盛尚纺织公司": "Li Ding Sheng Shang Fang Zhi Co., Ltd.",
+    "宝新科慧软件公司": "Bao Xin Ke Hui Ruan Jian Co., Ltd.",
+    "众课创信软件公司": "Zhong Ke Chuang Xin Ruan Jian Co., Ltd.",
+    "灿芯辉芯半导体公司": "Can Xin Hui Xin Semiconductor Co., Ltd.",
+    "瑞芯芯耀材料公司": "Rui Xin Xin Yao Materials Co., Ltd.",
+    "创玮耀耀电气公司": "Chuang Wei Yao Yao Dian Qi Co., Ltd.",
+    "美能电锦科技公司": "Mei Neng Dian Jin Technology Co., Ltd.",
+    "用丰信软网络公司": "Yong Feng Xin Ruan Network Co., Ltd.",
+    "金飞数软数据服务公司": "Jin Fei Shu Ruan Data Services Co., Ltd.",
+    "物丽汇锦零售公司": [
+      "Wu Li Hui Jin Retail Co., Ltd.",
+      "Wu Li Hui Jin Ling Shou Co., Ltd."
+    ],
+    "美能炫锦电气公司": "Mei Neng Xuan Jin Dian Qi Co., Ltd.",
+    "丽信耀悦电器公司": "Li Xin Yao Yue Dian Qi Co., Ltd.",
+    "海丽创耀家电公司": "Hai Li Chuang Yao Jia Dian Co., Ltd.",
+    "三夏泽能电力公司": "San Xia Ze Neng Dian Li Co., Ltd.",
+    "景能电热燃气公司": "Jing Neng Dian Re Ran Qi Co., Ltd.",
+    "以山泰泰医疗器械公司": "Yi Shan Tai Tai Medical Devices Co., Ltd.",
+    "光盛昌泽集团公司": "Guang Sheng Chang Ze Group Co., Ltd.",
+    "浪集云慧科技公司": "Lang Ji Yun Hui Technology Co., Ltd.",
+    "普各健辰药业股份公司": "Pu Ge Jian Chen Pharmaceutical Co., Ltd.",
+    "金湖地产建设开发公司": "Jin Hu Real Estate Construction Development Co., Ltd.",
+    "华城盛源综合开发公司": [
+      "Hua Cheng Sheng Yuan Integrated Development Co., Ltd.",
+      "Hua Cheng Sheng Yuan Zong He Kai Fa Co., Ltd."
+    ],
+    "龙河置锦置业公司": [
+      "Long He Zhi Jin Real Estate Co., Ltd.",
+      "Long He Zhi Jin Zhi Ye Co., Ltd."
+    ],
+    "十阳智光电器公司": "Shi Yang Zhi Guang Dian Qi Co., Ltd.",
+    "徐业智工科技公司": "Xu Ye Zhi Gong Technology Co., Ltd.",
+    "原通盛盛供应链公司": "Yuan Tong Sheng Sheng Gong Ying Lian Co., Ltd.",
+    "神轴务锦咨询公司": "Shen Zhou Wu Jin Zi Xun Co., Ltd.",
+    "众邮政运运港口公司": "Zhong You Zheng Yun Yun Port Co., Ltd.",
+    "用丰科联软件公司": "Yong Feng Ke Lian Software Co., Ltd.",
+    "海丽炫悦电气公司": "Hai Li Xuan Yue Electric Co., Ltd.",
+    "恒逸润恒科技股份公司": "Heng Yi Run Heng Technology Co., Ltd.",
+    "连机机锦机床公司": "Lian Ji Ji Jin Ji Chuang Co., Ltd.",
+    "平汝港通运物流公司": "Ping Ru Gang Tong Yun Wu Liu Co., Ltd.",
+    "卫星润锦科技股份公司": "Wei Xing Run Jin Ke Ji Co., Ltd.",
+    "高尹泽通批发公司": [
+      "Gao Yin Ze Tong Pi Fa Co., Ltd.",
+      "Gaoyin Zetong Wholesale Co., Ltd."
+    ],
+    "永惠泽汇批发公司": "Yong Hui Ze Hui Pi Fa Co., Ltd.",
+    "麻钢泰锦材料公司": [
+      "Ma Gang Tai Jin Cai Liao Co., Ltd.",
+      "Magang Taijin Materials Co., Ltd."
+    ],
+    "麻钢钢盛不锈钢公司": [
+      "Ma Gang Gang Sheng Bu Xiu Gang Co., Ltd.",
+      "Magang Gangsheng Stainless Steel Co., Ltd."
+    ],
+    "创玮耀盛电器公司": "Chuangwei Yaosheng Electric Co., Ltd.",
+    "丽信智创家电公司": "Lixin Zhichuang Home Appliances Co., Ltd.",
+    "碧园产华地产控股公司": "Biyuan Chanhua Real Estate Holdings Co., Ltd.",
+    "花润置锦建设开发公司": "Huarun Zhijin Construction Development Co., Ltd.",
+    "健名生康医疗器械公司": "Jian Ming Sheng Kang Yi Liao Qi Xie Co., Ltd.",
+    "康盛康瑞药业股份公司": "Kang Sheng Kang Rui Yao Ye Joint Stock Co., Ltd.",
+    "创玮锦智电器公司": "Chuang Wei Jin Zhi Electrical Appliances Co., Ltd.",
+    "龙河产置地产控股公司": "Long He Chan Zhi Di Chan Holdings Co., Ltd.",
+    "花盈泰盛财富管理公司": [
+      "Hua Ying Tai Sheng Wealth Management Co., Ltd.",
+      "Huaying Taisheng Wealth Management Co., Ltd.",
+      "Hua Ying Tai Sheng Cai Fu Management Co., Ltd."
+    ],
+    "用丰联创系统公司": [
+      "Yong Feng Lian Chuang Xi Tong Co., Ltd.",
+      "Yongfeng Lianchuang Systems Co., Ltd."
+    ],
+    "众白达贸批发公司": "Zhongbai Damao Wholesale Co., Ltd.",
+    "鲁西润恒化工公司": "Luxi Runheng Chemical Co., Ltd.",
+    "美能炫悦电气公司": [
+      "Meineng Xuanyue Electric Co., Ltd.",
+      "Mei Neng Xuan Yue Dian Qi Co., Ltd."
+    ],
+    "包铁源昌金属制品公司": [
+      "Baotie Yuanchang Metal Products Co., Ltd.",
+      "Bao Tie Yuan Chang Jin Shu Zhi Pin Co., Ltd."
+    ],
+    "保禾华昌建设开发公司": "Bao He Hua Chang Jian She Kai Fa Co., Ltd.",
+    "丽群通通电子商务公司": "Li Qun Tong Tong Dian Zi Shang Wu Co., Ltd.",
+    "亚玮泽智科技公司": "Ya Wei Ze Zhi Technology Co., Ltd.",
+    "创芯耀锐集成电路公司": "Chuang Xin Yao Rui Integrated Circuit Co., Ltd.",
+    "连机智盛机械公司": "Lian Ji Zhi Sheng Ji Xie Co., Ltd.",
+    "和联创航设备公司": "He Lian Chuang Hang She Bei Co., Ltd.",
+    "山拉达创智能装备公司": "Shan La Da Chuang Zhi Neng Zhuang Bei Co., Ltd.",
+    "三三工智科技公司": "Sansan Gongzhi Technology Co., Ltd.",
+    "杰杰达航设备公司": "Jiejie Dahang Equipment Co., Ltd.",
+    "众金冶冶资源公司": "Zhongjin Yeye Resources Co., Ltd.",
+    "海山味香餐饮管理公司": "Haishan Weixiang Catering Management Co., Ltd.",
+    "三松食锦调味品公司": "Sansong Shijin Condiment Co., Ltd.",
+    "美能电光家电公司": "Meineng Dianguang Home Appliances Co., Ltd.",
+    "丽信盛悦智能科技公司": "Lixin Shengyue Intelligent Technology Co., Ltd.",
+    "惠金锦瑞财富管理公司": [
+      "Huijin Jinrui Wealth Management Co., Ltd.",
+      "Huijin Jinrui Wealth Management Company",
+      "Hui Jin Jin Rui Cai Fu Management Co., Ltd."
+    ],
+    "众课智云数据服务公司": [
+      "Zhongke Zhiyun Data Services Co., Ltd.",
+      "Zhong Ke Zhi Yun Shu Ju Fu Wu Co., Ltd.",
+      "Zhongke Zhiyun Data Services Company"
+    ],
+    "长桥锦创科技公司": [
+      "Changqiao Jinchuang Technology Co., Ltd.",
+      "Zhang Qiao Jin Chuang Technology Co., Ltd.",
+      "Zhangqiao Jinchuang Technology Co., Ltd.",
+      "Changqiao Jinchuang Technology Company"
+    ],
+    "物丽汇达连锁公司": [
+      "Wuli Huida Chain Co., Ltd.",
+      "Wu Li Hui Da Chain Co., Ltd.",
+      "Wuli Huida Chain Company",
+      "Wu Li Hui Da Lian Suo Co., Ltd."
+    ],
+    "众课科数软件公司": [
+      "Zhong Ke Shu Ruan Software Co., Ltd.",
+      "Zhong Ke Ke Shu Software Co., Ltd.",
+      "Zhongke Keshu Software Company",
+      "Zhong Ke Ke Shu Ruan Jian Co., Ltd."
+    ],
+    "碧园盛华建设开发公司": [
+      "Bi Yuan Sheng Hua Jian She Kai Fa Co., Ltd.",
+      "Biyuan Shenghua Construction Development Co., Ltd."
+    ],
+    "碧园置泽城市发展公司": [
+      "Bi Yuan Zhi Ze Urban Development Co., Ltd.",
+      "Bi Yuan Zhi Ze Cheng Shi Development Co., Ltd.",
+      "Biyuan Zhize Urban Development Co., Ltd.",
+      "Biyuan Zhize Urban Development Company"
+    ],
+    "招业华昌房地产开发公司": [
+      "Zhao Ye Hua Chang Real Estate Development Co., Ltd.",
+      "Zhaoye Huachang Real Estate Development Co., Ltd.",
+      "Zhao Ye Hua Chang Fang Di Chan Kai Fa Co., Ltd.",
+      "Zhaoye Huachang Real Estate Development Company"
+    ],
+    "铜通泽鸿证券公司": [
+      "Tong Tong Ze Hong Securities Co., Ltd.",
+      "Tongtong Zehong Securities Co., Ltd.",
+      "Tong Tong Ze Hong Zheng Quan Co., Ltd."
+    ],
+    "爱健颐康复众心公司": [
+      "Aijian Yikang Fuzhongxin Co., Ltd.",
+      "Ai Jian Yi Kang Fu Zhong Xin Co., Ltd.",
+      "Aijian Yikang Fuzhongxin Company"
+    ],
+    "宝新慧慧网络公司": [
+      "Baoxin Huihui Network Co., Ltd.",
+      "Bao Xin Hui Hui Wang Luo Co., Ltd.",
+      "Baoxin Huihui Network Company"
+    ],
+    "众车远泽船舶公司": [
+      "Zhong Che Yuan Ze Shipbuilding Co., Ltd.",
+      "Zhongche Yuanze Shipbuilding Co., Ltd.",
+      "Zhongche Yuanze Shipbuilding Company"
+    ],
+    "花图文教在线教育公司": [
+      "Hua Tu Wen Jiao Zai Xian Jiao Yu Co., Ltd.",
+      "Huatu Wenjiao Online Education Co., Ltd."
+    ],
+    "金制鸿盛资产管理公司": [
+      "Jin Zhi Hong Sheng Zi Chan Guan Li Co., Ltd.",
+      "Jinzhi Hongsheng Asset Management Co., Ltd.",
+      "Jinzhi Hongsheng Asset Management Company",
+      "Jin Zhi Hong Sheng Zi Chan Management Co., Ltd."
+    ],
+    "健帆宁泽养老服务公司": [
+      "Jianfan Ningze Elderly Care Services Co., Ltd.",
+      "Jian Fan Ning Ze Yang Lao Fu Wu Co., Ltd."
+    ],
+    "一海昌锦商务公司": [
+      "Yihai Changjin Business Co., Ltd.",
+      "Yi Hai Chang Jin Shang Wu Co., Ltd.",
+      "Yihai Changjin Business Company"
+    ],
+    "浪集慧软科技公司": [
+      "Lang Ji Hui Ruan Technology Co., Ltd.",
+      "Langji Huiruan Technology Co., Ltd.",
+      "Langji Huiruan Technology Company"
+    ],
+    "花新源石新材料公司": [
+      "Hua Xin Yuan Shi New Materials Co., Ltd.",
+      "Huaxin Yuanshi New Materials Company",
+      "Hua Xin Yuan Shi Xin Cai Liao Co., Ltd."
+    ],
+    "碧园产锦不动产公司": [
+      "Bi Yuan Chan Jin Bu Dong Chan Co., Ltd.",
+      "Biyuan Chanjin Real Estate Company"
+    ],
+    "众海工昌锦建筑设计公司": [
+      "Zhonghai Gongchangjin Architectural Design Company",
+      "Zhong Hai Gong Chang Jin Jian Zhu She Ji Co., Ltd."
+    ],
+    "招业泽锦地产控股公司": [
+      "Zhao Ye Ze Jin Di Chan Holdings Co., Ltd.",
+      "Zhao Ye Ze Jin Real Estate Holdings Co., Ltd."
+    ],
+    "瑞行健康制药公司": [
+      "Rui Xing Jian Kang Zhi Yao Co., Ltd.",
+      "Ruixing Health Pharmaceutical Company"
+    ],
+    "北控泽净水务公司": [
+      "Beikong Zejing Water Company",
+      "Bei Kong Ze Jing Water Co., Ltd."
+    ],
+    "恒丽科智软件公司": [
+      "Hengli Kezhi Software Company",
+      "Heng Li Ke Zhi Ruan Jian Co., Ltd."
+    ],
+    "十阳锦锦电器公司": [
+      "Shiyang Jinjin Electrical Appliances Company",
+      "Shi Yang Jin Jin Electrical Appliances Co., Ltd."
+    ],
+    "星酷文工艺美术品公司": [
+      "Xingkuwen Arts and Crafts Company",
+      "Xing Ku Wen Gong Yi Mei Shu Pin Co., Ltd."
+    ],
+    "新花源通连锁公司": "Xin Hua Yuan Tong Lian Suo Co., Ltd.",
+    "花电能锦水电公司": [
+      "Hua Dian Neng Jin Shui Dian Co., Ltd.",
+      "Hua Dian Neng Jin Hydropower Co., Ltd."
+    ],
+    "花能泽泽新能源公司": "Hua Neng Ze Ze Xin Neng Yuan Co., Ltd.",
+    "航发铁船航空科技公司": "Hang Fa Tie Chuan Hang Kong Technology Co., Ltd.",
+    "润会数科科技公司": "Run Hui Shu Ke Technology Co., Ltd.",
+    "众防昌达重工公司": "Zhong Fang Chang Da Zhong Gong Co., Ltd.",
+    "宝新智智系统公司": "Bao Xin Zhi Zhi Xi Tong Co., Ltd.",
+    "大族锦精设备公司": "Da Zu Jin Jing She Bei Co., Ltd.",
+    "锡份冶锦金属公司": "Xi Fen Ye Jin Jin Shu Co., Ltd.",
+    "绿泰洁循环保科技公司": "Lv Tai Jie Xun Huan Bao Technology Co., Ltd.",
+    "丰火创泽网络设备公司": "Feng Huo Chuang Ze Wang Luo She Bei Co., Ltd.",
+    "环丘泰锦智能电气公司": "Huan Qiu Tai Jin Zhi Neng Dian Qi Co., Ltd.",
+    "绿山置锦房地产开发公司": "Lv Shan Zhi Jin Real Estate Development Co., Ltd.",
+    "瑞芯耀澜集成电路公司": "Rui Xin Yao Lan Ji Cheng Dian Lu Co., Ltd.",
+    "晶芯锐辉微电子公司": "Jing Xin Rui Hui Wei Dian Zi Co., Ltd.",
+    "药石元泽生物医药公司": "Yao Shi Yuan Ze Sheng Wu Yi Yao Co., Ltd.",
+    "众集昌源钢铁公司": "Zhong Ji Chang Yuan Gang Tie Co., Ltd.",
+    "浪集联创信息技术公司": "Lang Ji Lian Chuang Xin Xi Ji Shu Co., Ltd.",
+    "潞安富昌煤炭公司": "Lu An Fu Chang Mei Tan Co., Ltd.",
+    "恒丽云创信息技术公司": "Heng Li Yun Chuang Xin Xi Ji Shu Co., Ltd.",
+    "万会盛置建设开发公司": "Wan Hui Sheng Zhi Construction Development Co., Ltd.",
+    "航发远锦航空科技公司": "Hang Fa Yuan Jin Hang Kong Technology Co., Ltd.",
+    "华鲁荣荣化学公司": "Hua Lu Rong Rong Hua Xue Co., Ltd.",
+    "三三工机科技公司": "San San Gong Ji Technology Co., Ltd.",
+    "众聚悦饮食品公司": "Zhong Ju Yue Yin Shi Pin Co., Ltd.",
+    "华城锦锦综合开发公司": "Hua Cheng Jin Jin Zong He Kai Fa Co., Ltd.",
+    "绿山产锦置业公司": "Lv Shan Chan Jin Zhi Ye Co., Ltd.",
+    "花新泽昌新材料公司": "Hua Xin Ze Chang Xin Cai Liao Co., Ltd.",
+    "万会锦盛房地产开发公司": [
+      "Wan Hui Jin Sheng Real Estate Development Co., Ltd.",
+      "Wan Hui Jin Sheng Fang Di Chan Kai Fa Co., Ltd."
+    ],
+    "包金金昌铜业公司": "Bao Jin Jin Chang Tong Ye Co., Ltd.",
+    "众邮政达锦运输公司": "Zhong You Zheng Da Jin Yun Shu Co., Ltd.",
+    "环星锦雅服饰公司": "Huan Xing Jin Ya Apparel Co., Ltd.",
+    "豪美公司": "Haomei Company"
+  },
+  "policy_items": {
+    "原材料工业\"三品\"实施方案": "Implementation Plan for \"Three Products\" in Raw Materials Industry",
+    "支持技工强省建设若干政策": "Several Policies on Supporting the Construction of a Strong Province of Skilled Workers",
+    "安徽省人民政府关于印发支持技工强省建设若干政策的通知": "Notice of Anhui Provincial People's Government on Several Policies Supporting the Construction of a Strong Province of Skilled Workers",
+    "广东省人民政府办公厅关于印发广东省进一步促进工业经济平稳增长若干措施的通知": "Notice of the General Office of Guangdong Provincial People's Government on Printing and Distributing Several Measures of Guangdong Province for Further Promoting Steady Growth of Industrial Economy",
+    "关于组织申报生物医药产业优秀青年人才首套房购买补贴的通知": "Notice on Organizing Applications for First Home Purchase Subsidies for Outstanding Young Talents in the Biomedicine Industry",
+    "东数西算": "Eastern Data Western Computing",
+    "广东省人民政府办公厅关于印发广东省进一步 促进工业经济平稳增长若干措施的通知": "Notice of Guangdong Provincial Development and Reform Commission on Issuing the Implementation Plan for Building Guangdong's Modern Circulation System during the 14th Five-Year Plan",
+    "广东省发展改革委关于印发《广东省“十四五”现代流通体系建设实施方案》": "Notice of the General Office of the People's Government of Guangdong Province on Issuing Several Measures to Further Promote Stable Growth of the Industrial Economy in Guangdong Province",
+    "关于上海市浦东新区有关研发机构适用进口税收政策资格认定事项的通知": "Notice on Qualification Recognition Matters for Relevant R&D Institutions in Pudong New Area of Shanghai Applicable to Import Tax Policies",
+    "上海市人民政府办公厅关于印发培育“元宇宙”新赛道行动方案的通知": "Notice of the General Office of the Shanghai Municipal People's Government on Issuing the Action Plan for Cultivating the New Metaverse Track",
+    "黑龙江省人民政府办公厅关于印发黑龙江省科技振兴行动计划（2022—2026年）的通知": "Notice of the General Office of the People's Government of Heilongjiang Province on Issuing the Heilongjiang Province Science and Technology Revitalization Action Plan (2022-2026)",
+    "黑龙江省支持大型国际合作交流活动对接引进优秀人才经费补助实施细则（试行）": "Detailed Rules for the Implementation of Funding Subsidies for Supporting Large-scale International Cooperation and Exchange Activities to Connect and Introduce Outstanding Talents in Heilongjiang Province (Trial)"
+  }
+}

assets/database/enterprise/company_core.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e7fadec7a8833ef4c5736d638c62bd1033247e50a47afb2fa36b0af81fa6a0f
+size 27411510

assets/database/enterprise/company_operation_status.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d30f8659da51de379500f0314d5d6322d534267e92377ca9eea47336e9fff94
+size 34717208

assets/database/enterprise/company_operation_status_detail.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:016e39c51abb7f940506d2855bbc0419a6ba50dbbe14e62684285d6c7985d569
+size 70022647

assets/database/enterprise/company_operation_yearly_status.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df4aa7baff8cea1d9eadec7b6650d57a827ea0520cb3fe636d708afadcb0fd15
+size 65472033

assets/database/enterprise/company_profile.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecd68737f4f453746c3271b146ef5f6b6e741536aab4fc8ae2a32be89085e3c2
+size 1783742

assets/database/enterprise/company_profile_as.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14488f5e36b9eec554bc1776513135206dc83b3fc834441bc7eff4ca5171a3cc
+size 242816

assets/database/enterprise/company_profile_eu.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:448bd2077692933fb41398dff37f7767c2cac726d852bf54d5ea9ead2b1c4a8e
+size 228500

assets/database/enterprise/company_profile_na.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83323ec3a56cb36abbb86263c95ee39e0ad5b2bff5101e90d95a69d5efe158cf
+size 152112

assets/database/enterprise/company_profile_oc.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f174f6be68dfd8ccfabfc50fed443faa3ffc339b23012a6e54a262b72d19981
+size 79246

assets/database/industry/national_industry_status.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:723ce91a24f67b5a3d032a297d40417a1df32c2d584c8215047de92f7c204fd6
+size 1472926

assets/database/industry/national_industry_status_detail.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f040fb21b4208bc280acc3bb6c81feb286bbd63150a30ca22677ed802d529b7
+size 3631383

assets/database/industry/national_industry_yearly_status.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6a2f88f64e19168d8adb70b0edd9dbeebd8df44c4db472c5b23af7a8b0d0f85
+size 55980

assets/database/industry/regional_industry_status.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6314f5198f67c8ebf90e03f891f042d8244a5cee5538bd4731678820b289d7e
+size 22504920

assets/database/industry/regional_industry_status_detail.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a08132df31975bcfbb56ea2d2a3b79dc3b052b97d0f88b63ebe1542a0d6f822d
+size 287796

assets/database/industry/regional_industry_yearly_status.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab96eb7f4ca90cf54cd75515beb6031ed72ff26d2df8d3264a969b1c2d8aab41
+size 289546

assets/database/internal_metrics.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f584fb7cacf19fcdc50e59b398fc65ad041d9651860d9cfe1d7afb277dbeb9dc
+size 2238

assets/database/policy/policy_release_status.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a784a9301e6a8feb63d7b5e2fae7da0f4a3fee3d1871703cf1560d36163184ef
+size 1109101

assets/database/policy/policy_resource.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:041c1e89ed2a24164075e92c9b8e7f74490042ef52dcfbd2c0e8f5c17764a029
+size 20833614

assets/qa_gold/comprehensive_decision/easy001.json

@@ -0,0 +1,22 @@
+{
+  "id": "easy001",
+  "question": "In which province is the enterprise with the highest R&D investment ratio nationwide in 2022 located?",
+  "guidelines": "The answer must be \"yes\" or \"no\". Output only the answer, do not add any explanatory text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": "Jiangsu Province",
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "easy",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter all records with year=2022 from company_operation_status.csv, and extract the enterprise name and R&D investment ratio fields.",
+    "Filter enterprise records with non-null R&D investment ratio, sort by R&D investment ratio in descending order, and identify the enterprise with the highest R&D investment ratio as \"Yaoshi Shenkang Medical Equipment Company\" with an R&D investment ratio of 934642.80%.",
+    "Look up the record for enterprise name \"Yaoshi Shenkang Medical Equipment Company\" in company_profile.csv, extract the province field, and obtain the province where the enterprise is located as \"Jiangsu Province\"."
+  ],
+  "steps_num": 3,
+  "milestone": {
+    "Enterprise with highest R&D investment ratio": "Yaoshi Shenkang Medical Equipment Company",
+    "R&D investment ratio (%)": 934642.8,
+    "Province of location": "Jiangsu Province"
+  }
+}

assets/qa_gold/comprehensive_decision/easy002.json

@@ -0,0 +1,36 @@
+{
+  "id": "easy002",
+  "question": "In 2022, nationwide, how is the chemical raw materials and chemical products manufacturing industry ranked by asset scale? Please list the top five provinces.",
+  "guidelines": "Answer format: [Province A, Province B, ...]. Output only province names, do not add any other explanatory text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": [
+    "Shandong Province",
+    "Zhejiang Province",
+    "Jiangsu Province",
+    "Shanghai",
+    "Guangdong Province"
+  ],
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "easy",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter all records with industry=\"Chemical Raw Materials and Chemical Products Manufacturing\" from regional_industry_status.csv, and extract the province and total assets fields.",
+    "Sort all provincial data by total assets in descending order to determine the asset scale ranking of each province. Extract the top five provinces and their total assets: Shandong Province (544109686731.85), Zhejiang Province (401644798867.80), Jiangsu Province (250743006622.33), Shanghai (230472528900.23), Guangdong Province (176926240169.03)."
+  ],
+  "steps_num": 2,
+  "milestone": {
+    "Total assets in Shandong Province (yuan)": 544109686731.85,
+    "Total assets in Zhejiang Province (yuan)": 401644798867.8,
+    "Total assets in Jiangsu Province (yuan)": 250743006622.33,
+    "Total assets in Shanghai (yuan)": 230472528900.23,
+    "Total assets in Guangdong Province (yuan)": 176926240169.03,
+    "Top five provinces": [
+      "Shandong Province",
+      "Zhejiang Province",
+      "Jiangsu Province",
+      "Shanghai",
+      "Guangdong Province"
+    ]
+  }
+}

assets/qa_gold/comprehensive_decision/easy003.json

@@ -0,0 +1,23 @@
+{
+  "id": "easy003",
+  "question": "In 2022, in the Qilu region (Shandong Province), how many policies support Zhongbai Jinmao Chain Company in its industry?",
+  "guidelines": "The answer must be an exact number. Output only the number, do not add units, commas, or any textual explanation. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 2,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "easy",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Search for records of \"Zhongbai Jinmao Chain Company\" in company_profile.csv, extract the industry field, and determine that its industry is \"Wholesale and Retail Trade\".",
+    "Filter from policy_release_status.csv for all policy records with province=\"Shandong Province\", and policies where the industry field contains \"Wholesale and Retail Trade\". Found 2 policies: 1 from \"Local Policy - Shandong Provincial People's Government General Office Policy Count\" and 1 from \"Local Policy - Shandong Provincial Development and Reform Commission Policy Count\".",
+    "Count the number of policies meeting the criteria, which is 2."
+  ],
+  "steps_num": 3,
+  "milestone": {
+    "Industry of Zhongbai Jinmao Chain Company": "Wholesale and Retail Trade",
+    "Shandong Province Local Policy - Shandong Provincial People's Government General Office Policy Count": 1,
+    "Shandong Province Local Policy - Shandong Provincial Development and Reform Commission Policy Count": 1,
+    "Number of policies in Shandong Province related to Wholesale and Retail Trade": 2
+  }
+}

assets/qa_gold/comprehensive_decision/easy004.json

@@ -0,0 +1,36 @@
+{
+  "id": "easy004",
+  "question": "In 2022, nationwide, what is the ranking of provinces by asset size in the Information Transmission, Software and Information Technology Services industry? Please list the top five provinces.",
+  "guidelines": "Answer format: [Province A, Province B, ...]. Output only the province ranking, do not add any other explanatory text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": [
+    "Beijing",
+    "Zhejiang Province",
+    "Guangdong Province",
+    "Shanghai",
+    "Jiangsu Province"
+  ],
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "easy",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter from regional_industry_status.csv all records with industry=\"Information Transmission, Software and Information Technology Services\", and extract the province and total assets fields.",
+    "Sort all provincial data in descending order by total assets to determine the asset size ranking of each province. Extract the top five provinces and their total assets: Beijing (10297490896006.5), Zhejiang Province (4115693929492.25), Guangdong Province (2262247330030.01), Shanghai (1282711003966.55), Jiangsu Province (177568006242.47)."
+  ],
+  "steps_num": 2,
+  "milestone": {
+    "Total assets of Beijing (CNY)": 10297490896006.5,
+    "Total assets of Zhejiang Province (CNY)": 4115693929492.25,
+    "Total assets of Guangdong Province (CNY)": 2262247330030.01,
+    "Total assets of Shanghai (CNY)": 1282711003966.55,
+    "Total assets of Jiangsu Province (CNY)": 177568006242.47,
+    "Top five provinces by ranking": [
+      "Beijing",
+      "Zhejiang Province",
+      "Guangdong Province",
+      "Shanghai",
+      "Jiangsu Province"
+    ]
+  }
+}

assets/qa_gold/comprehensive_decision/easy005.json

@@ -0,0 +1,37 @@
+{
+  "id": "easy005",
+  "question": "In 2022, nationwide, what is the ranking of provinces by profitability in the Real Estate industry? Please list the top five provinces.",
+  "guidelines": "Answer format: [Province A, Province B, ...]. Output only the province ranking, do not add any other explanatory text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": [
+    "Hong Kong SAR",
+    "Guangdong Province",
+    "Zhejiang Province",
+    "Beijing",
+    "Jilin Province"
+  ],
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "easy",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter from regional_industry_status.csv all records with industry=\"Real Estate\", extract province and total net profit amount fields, finding data for 34 provinces/regions in the Real Estate industry.",
+    "Filter province records where total net profit amount is not null, totaling 34 valid records. Sort all provincial data in descending order by total net profit amount to determine the profitability ranking of each province.",
+    "Extract the top five provinces and their total net profit amounts: Hong Kong SAR (86497465420.52 CNY), Guangdong Province (70559018502.22 CNY), Zhejiang Province (12297928184.06 CNY), Beijing (8975618268.55 CNY), Jilin Province (675521026.00 CNY)."
+  ],
+  "steps_num": 3,
+  "milestone": {
+    "Total net profit amount of Hong Kong SAR (CNY)": 86497465420.52,
+    "Total net profit amount of Guangdong Province (CNY)": 70559018502.22,
+    "Total net profit amount of Zhejiang Province (CNY)": 12297928184.06,
+    "Total net profit amount of Beijing (CNY)": 8975618268.55,
+    "Total net profit amount of Jilin Province (CNY)": 675521026.0,
+    "Top five provinces by ranking": [
+      "Hong Kong SAR",
+      "Guangdong Province",
+      "Zhejiang Province",
+      "Beijing",
+      "Jilin Province"
+    ]
+  }
+}

assets/qa_gold/comprehensive_decision/easy006.json

@@ -0,0 +1,21 @@
+{
+  "id": "easy006",
+  "question": "In 2022, what is Sichuan Province's national ranking by average R&D investment in the Information Transmission, Software and Information Technology Services industry?",
+  "guidelines": "The answer must be an exact number representing the ranking. Output only the number, do not add units, commas, or any textual explanation. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 10,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "easy",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter from regional_industry_status.csv all records with industry=\"Information Transmission, Software and Information Technology Services\", finding data for 34 provinces/regions.",
+    "Extract the \"mean R&D investment amount\" field for each province to obtain mean R&D investment data for all provinces. Sort all provinces by mean R&D investment amount in descending order to determine the R&D investment ranking of each province.",
+    "Locate Sichuan Province's position in the sorted list. Sichuan Province's mean R&D investment amount is 147274301.44 CNY, and its ranking is 10th."
+  ],
+  "steps_num": 3,
+  "milestone": {
+    "Mean R&D investment amount of Sichuan Province (CNY)": 147274301.44,
+    "Sichuan Province ranking": 10
+  }
+}

assets/qa_gold/comprehensive_decision/hard001.json

@@ -0,0 +1,31 @@
+{
+  "id": "hard001",
+  "question": "In 2022, a strategic consulting firm was commissioned by a provincial government to quantitatively rank the comprehensive attractiveness of pharmaceutical manufacturing across provinces, in order to identify priority target regions for attracting leading enterprises. The company designed a four-dimensional weighted scoring system: four original indicators—enterprise agglomeration level (weight 30%), R&D expenditure as a share of revenue (weight 30%), regional policy coverage intensity (weight 20%), and R&D human resource penetration rate (weight 20%)—were normalized (min-max) and then weighted to produce a composite score. Among these, agglomeration level is measured by the proportion of enterprises in each province to the national total in pharmaceutical manufacturing; policy intensity is measured by the ratio of relevant policy items in each province to the total number of relevant policies nationwide; and human resource penetration rate is the total number of R&D personnel in each province divided by total employees. What is the specific composite score value of the province with the highest weighted composite score after normalization across provinces?",
+  "guidelines": "The answer should be a numerical value with 2 decimal places. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 0.92,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter records with industry=\"Pharmaceutical Manufacturing\" from regional_industry_status.csv, extract province, total enterprises, total R&D expenditure, total operating revenue, total R&D personnel, and total employees; 34 provincial records were found.",
+    "Filter records with industry=\"Pharmaceutical Manufacturing\" from national_industry_status.csv to obtain the national total of 449 pharmaceutical manufacturing enterprises.",
+    "Filter policy records from policy_release_status.csv where the industry field contains \"Pharmaceutical Manufacturing\"; 80 relevant policies were found across 22 provinces. Group by province to count policy numbers per province.",
+    "Calculate four original indicators for each province: industry agglomeration = total enterprises/449, R&D intensity = total R&D expenditure/total operating revenue, policy support = province policy count/80, talent density = total R&D personnel/total employees. Filter to 16 valid provinces with all four indicators non-null.",
+    "Apply min-max normalization to each of the four indicators across the 16 valid provinces: normalized value = (original value - min)/(max - min).",
+    "Calculate composite score = normalized industry agglomeration × 0.3 + normalized R&D intensity × 0.3 + normalized policy support × 0.2 + normalized talent density × 0.2.",
+    "Sort by composite score in descending order. Shanghai has the highest composite score, with industry agglomeration 0.1203, R&D intensity 0.2548, policy support 0.1375, talent density 0.1620, and composite score = 0.9160."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "National total pharmaceutical manufacturing enterprises": 449.0,
+    "National total pharmaceutical manufacturing-related policies": 80,
+    "Number of valid provinces": 16,
+    "Shanghai industry agglomeration": 0.1203,
+    "Shanghai R&D intensity": 0.2548,
+    "Shanghai policy support": 0.1375,
+    "Shanghai talent density": 0.162,
+    "Shanghai composite score": 0.916
+  }
+}

assets/qa_gold/comprehensive_decision/hard002.json

@@ -0,0 +1,28 @@
+{
+  "id": "hard002",
+  "question": "In 2022, conduct a quantitative assessment of the investment value of the semiconductor industry across provinces. The evaluation framework requires incorporating three dimensions: first, industry scale (40% weight), measured by the inter-provincial rank percentile of total operating revenue in each province; second, profitability quality (30% weight), reflected by the inter-provincial rank percentile of operating profit margin (total operating profit divided by total operating revenue) in each province; third, technology output intensity (30% weight), measured by the inter-provincial rank percentile of the ratio of total patent applications to R&D expenditure (converted to 100 million yuan) in each province. The rank percentile for each indicator is calculated by sorting values from low to high, using the formula (rank - 1) / (total number of provinces - 1). Note that only provinces with complete data for all three indicators are included in the calculation. Under this weighted scoring system, what is the final score of the province ranked first in comprehensive investment value?",
+  "guidelines": "The answer should be a numerical value with 2 decimal places. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 0.67,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter records with industry=\"Semiconductor Industry\" from regional_industry_status.csv, extract province, total operating revenue, total operating profit, and total R&D expenditure; 34 records were found.",
+    "Filter enterprises with industry=\"Semiconductor Industry\" from company_profile.csv, extract company name, bmCode, and province; 172 enterprises were found.",
+    "Join with company_operation_status.csv via bmCode to extract annual domestic invention patent applications. Filter 149 valid records with non-null values, group by province to sum annual domestic invention patent applications; 22 provinces have patent data.",
+    "Inner join regional data with enterprise-level patent summary by province, filter records with total operating revenue > 0 and total R&D expenditure > 0; 13 valid provinces. Calculate three original indicators: industry scale = total operating revenue, profitability = total operating profit / total operating revenue, innovation output = total annual domestic invention patent applications / total R&D expenditure (in 100 million yuan).",
+    "Rank each indicator by value from low to high (rank method = min), calculate rank percentile = (rank - 1) / (13 - 1).",
+    "Calculate investment value composite score = industry scale rank percentile × 0.4 + profitability rank percentile × 0.3 + innovation output rank percentile × 0.3.",
+    "Sort by composite score in descending order. Zhejiang Province has the highest composite score, with industry scale rank percentile 0.6667, profitability rank percentile 0.5000, innovation output rank percentile 0.8333, and composite score = 0.6667."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Number of valid provinces": 13,
+    "Zhejiang Province industry scale rank percentile": 0.6667,
+    "Zhejiang Province profitability rank percentile": 0.5,
+    "Zhejiang Province innovation output rank percentile": 0.8333,
+    "Zhejiang Province composite score": 0.6667
+  }
+}

assets/qa_gold/comprehensive_decision/hard003.json

@@ -0,0 +1,36 @@
+{
+  "id": "hard003",
+  "question": "In 2022, an automotive manufacturing enterprise commissioned a third-party institution to score and rate the industrial supporting capacity of each province before selecting a site for a new plant. The scoring rules are as follows: first, rank provinces in descending order by the number of government policies related to automotive manufacturing, and take the top five provinces by policy count as the candidate pool; then, within the candidate provinces, calculate the industrial supporting composite index, which is a weighted combination of three components—upstream and downstream supply chain density (weight 0.4), local labor reserve (weight 0.3), and government subsidy intensity per enterprise (weight 0.3). Supply chain density is defined as the ratio of total automotive manufacturing enterprises in the province to the national total in the industry; labor reserve is defined as the ratio of total industry employees in the province to the national total in the industry; subsidy intensity is defined as total government rewards and subsidies for automotive manufacturing in the province divided by the number of enterprises in the province (subsidy intensity must be normalized across all provinces before being used in the formula). Among the top five provinces by policy ranking, what is the composite index value of the province with the highest industrial supporting composite index?",
+  "guidelines": "Answer format: numerical value (4 decimal places). Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 0.3187,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter policy records from policy_release_status.csv where industry field contains \"Automotive Manufacturing\"; 69 records found. Group by province to count policy numbers per province (excluding \"National\" level), sort by policy count descending; top 5 provinces by policy support are ['Guangdong Province', 'Shanghai', 'Hunan Province', 'Sichuan Province', 'Chongqing'].",
+    "Filter records with industry=\"Automotive Manufacturing\" from regional_industry_status.csv, extract province, total enterprises, total employees, and total government rewards and subsidies; 34 records found.",
+    "Filter records with industry=\"Automotive Manufacturing\" from national_industry_status.csv to obtain national total of 230 automotive manufacturing enterprises and 3,254,510 total employees.",
+    "Calculate three original indicators for each province: upstream-downstream enterprise density = total enterprises/230, talent reserve = total employees/3,254,510, subsidy intensity = total government rewards and subsidies/total enterprises.",
+    "Apply min-max normalization to subsidy intensity: normalized value = (subsidy intensity - 5468453.88)/(476294284.71 - 5468453.88).",
+    "Calculate industrial supporting composite index = upstream-downstream enterprise density × 0.4 + talent reserve × 0.3 + normalized subsidy intensity × 0.3.",
+    "Filter among top 5 policy provinces ['Guangdong Province', 'Shanghai', 'Hunan Province', 'Sichuan Province', 'Chongqing'] (Chongqing excluded due to missing data for composite index calculation). Guangdong Province has the highest composite index: upstream-downstream enterprise density 0.1174, talent reserve 0.4499, normalized subsidy intensity 0.4558, composite index = 0.3187."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "National total automotive manufacturing enterprises": 230.0,
+    "National total automotive manufacturing employees": 3254510.0,
+    "Top 5 provinces by policy": [
+      "Guangdong Province",
+      "Shanghai",
+      "Hunan Province",
+      "Sichuan Province",
+      "Chongqing"
+    ],
+    "Guangdong Province upstream-downstream enterprise density": 0.1174,
+    "Guangdong Province talent reserve": 0.4499,
+    "Guangdong Province normalized subsidy intensity": 0.4558,
+    "Guangdong Province composite index": 0.3187
+  }
+}

assets/qa_gold/comprehensive_decision/hard004.json

@@ -0,0 +1,31 @@
+{
+  "id": "hard004",
+  "question": "In 2022, a provincial development and reform commission, when reviewing the effectiveness of fiscal subsidies for the chemical raw materials and chemical products manufacturing industry, needed to identify enterprises with misallocated subsidy resources. Specifically, analysts must first define the scope: only examine enterprises located in provinces that have policy entries for \"Chemical Raw Materials and Chemical Products Manufacturing\" in the policy release status data; then use the industry-wide median of government subsidies and the median operating profit margin (profit margin = operating profit ÷ operating revenue × 100%) as dual thresholds to identify \"capital misallocation\" enterprises—those that simultaneously have \"subsidy amount above the industry median\" but \"profit margin below the industry median\". Among the valid enterprises in the policy-covered provinces, what is the proportion of capital misallocation enterprises as a percentage of total valid enterprises in those provinces (express the result as a percentage with 2 decimal places, without the % symbol)?",
+  "guidelines": "The answer should be a percentage value with 2 decimal places. Output only the number, without the % symbol or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 23.18,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter policy records from policy_release_status.csv where industry field contains \"Chemical Raw Materials and Chemical Products Manufacturing\"; 60 records found. Extract province field, remove nulls and exclude \"National\" level; 23 policy-covered provinces: ['Shanghai', 'Yunnan Province', 'Inner Mongolia Autonomous Region', 'Sichuan Province', 'Ningxia Hui Autonomous Region', 'Anhui Province', 'Shandong Province', 'Shanxi Province', 'Guangdong Province', 'Guangxi Zhuang Autonomous Region', 'Xinjiang Uygur Autonomous Region', 'Jiangxi Province', 'Hebei Province', 'Henan Province', 'Hainan Province', 'Hubei Province', 'Hunan Province', 'Gansu Province', 'Fujian Province', 'Guizhou Province', 'Liaoning Province', 'Shaanxi Province', 'Heilongjiang Province'].",
+    "Filter enterprise records with industry=\"Chemical Raw Materials and Chemical Products Manufacturing\" from company_profile.csv, extract company name, bmCode, and province; 364 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract government rewards and subsidies, operating profit, and operating revenue; 364 records after merge.",
+    "Filter valid samples with non-null values for government rewards and subsidies, operating profit, and operating revenue; 362 enterprises.",
+    "Calculate industry-wide median benchmarks for valid enterprises: operating profit margin = operating profit/operating revenue × 100%; median government subsidy is 10,019,029.08 yuan, median operating profit margin is 10.00%.",
+    "Filter valid enterprises whose province is in the policy-covered province list; 233 enterprises.",
+    "Among the 233 valid enterprises in policy-covered provinces, filter \"high subsidy, low output\" enterprises with government subsidy > 10,019,029.08 and operating profit margin < 10.00%; 54 enterprises, proportion = 54/233 × 100% = 23.18%."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Number of policy-covered provinces": 23,
+    "Total chemical enterprises": 364,
+    "Number of valid samples": 362,
+    "Median government subsidy (yuan)": 10019029.08,
+    "Median operating profit margin (%)": 10.0,
+    "Valid enterprises in policy-covered provinces": 233,
+    "High subsidy low output enterprises": 54,
+    "Proportion (%)": 23.18
+  }
+}

assets/qa_gold/comprehensive_decision/hard005.json

@@ -0,0 +1,30 @@
+{
+  "id": "hard005",
+  "question": "In 2022, for the information transmission, software and information technology services industry, an industry research institute sought to obtain a policy-adjusted comprehensive innovation efficiency indicator by superimposing the incentive effect of local policy support on top of raw innovation efficiency. The calculation logic is as follows: first, exclude from enterprise microdata any samples with missing R&D expenditure or annual domestic invention patent grants; for the remaining valid enterprises, aggregate by province and calculate the ratio of total invention patent grants to total R&D expenditure (converted to 100 million yuan) for each province as the province's raw innovation efficiency benchmark; then use the proportion of policy items in that province out of all information technology policies as the policy support coefficient, and multiply the raw efficiency benchmark by (1 plus the policy support coefficient) to obtain the final policy-adjusted innovation efficiency. Among all provinces with data, what is the specific value of the province with the highest adjusted efficiency?",
+  "guidelines": "The answer should be a numerical value with 2 decimal places. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 63.74,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter policy records from policy_release_status.csv where industry field contains \"Information Transmission, Software and Information Technology Services\"; 206 records found. Group by province to count policy numbers per province.",
+    "Filter enterprise records with industry=\"Information Transmission, Software and Information Technology Services\" from company_profile.csv, extract company name, bmCode, and province; 644 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract R&D expenditure and annual domestic invention patent grants; 644 records after merge.",
+    "Filter valid enterprises with non-null values for both R&D expenditure and annual domestic invention patent grants; 432 enterprises.",
+    "Group by province to sum R&D expenditure and annual domestic invention patent grants; 28 provinces have valid data.",
+    "Convert total R&D expenditure per province to 100 million yuan, calculate raw innovation efficiency = total patent grants / total R&D expenditure (100 million yuan). Merge with policy data, calculate policy support coefficient = province policy count / 206.",
+    "Calculate policy-adjusted innovation efficiency = raw innovation efficiency × (1 + policy support coefficient), sort by adjusted efficiency descending. Hong Kong Special Administrative Region has the highest: raw efficiency 63.7360, policy support coefficient 0.0000, adjusted efficiency = 63.7360."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Total information technology-related policies": 206,
+    "Total information technology services enterprises": 644,
+    "Number of valid enterprises": 432,
+    "Number of valid provinces": 28,
+    "Hong Kong SAR raw innovation efficiency": 63.736,
+    "Hong Kong SAR policy support coefficient": 0.0,
+    "Hong Kong SAR adjusted innovation efficiency": 63.74
+  }
+}

assets/qa_gold/comprehensive_decision/hard006.json

@@ -0,0 +1,32 @@
+{
+  "id": "hard006",
+  "question": "In 2022, to measure the impact of different ownership backgrounds on the operating performance of specialized equipment manufacturing enterprises, an analysis team compared each enterprise's return on equity (ROE) level with the industry-wide return level in its province to calculate \"excess ROE\" as a relative performance indicator. Specifically: enterprise ROE is calculated as net profit divided by net assets (total assets minus total liabilities) multiplied by 100%; provincial industry benchmark ROE is extracted from provincial industry summary tables, calculated as total industry net profit divided by total industry net assets (total assets minus total liabilities) multiplied by 100%; each enterprise's excess ROE is the difference between its own ROE and its province's benchmark ROE. After grouping by ownership type, which ownership category has the highest mean excess ROE among enterprises? What is that mean value in percentage points?",
+  "guidelines": "The answer should be a percentage value with 2 decimal places. Output only the number, without the % symbol or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": [
+    "Collective Enterprise",
+    5.17
+  ],
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter enterprise records with industry=\"Specialized Equipment Manufacturing\" from company_profile.csv, extract company name, bmCode, ownership, and province; 447 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract net profit, total assets, total liabilities, and operating revenue; 447 records after merge.",
+    "Filter records with industry=\"Specialized Equipment Manufacturing\" from regional_industry_status.csv, calculate provincial industry average ROE = total net profit / (total assets - total liabilities) × 100%; 15 valid provinces.",
+    "Filter valid enterprises with total assets > total liabilities and non-null operating revenue; 444 enterprises.",
+    "Calculate net assets = total assets - total liabilities for each enterprise, then ROE = net profit / net assets × 100%.",
+    "Inner join enterprise data with provincial industry average ROE by province; 389 enterprises matched. Calculate excess ROE = enterprise ROE - provincial industry average ROE for each enterprise.",
+    "Group by ownership to calculate mean excess ROE for each ownership type; 6 ownership types. Collective enterprises (2 enterprises) have the highest mean excess ROE = 5.17%."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Total specialized equipment manufacturing enterprises": 447,
+    "Number of valid enterprises": 444,
+    "Enterprises matched with provincial data": 389,
+    "Number of ownership types": 6,
+    "Collective enterprise count": 2,
+    "Collective enterprise mean excess ROE (%)": 5.17
+  }
+}

assets/qa_gold/comprehensive_decision/hard007.json

@@ -0,0 +1,39 @@
+{
+  "id": "hard007",
+  "question": "In 2022, when a research institution was reviewing the implementation effectiveness of provincial industrial policies in the food and beverage industry, it found that although some provinces had issued many support policies, the profitability of enterprises within their jurisdictions was not ideal. To identify such \"policy-heavy, low-return\" provinces, the institution planned to analyze separately those provinces with a relatively large number of policies (including national-level policies, totaling 3 or more): sum the operating profit amounts of all food and beverage industry enterprises in these provinces and divide by the sum of operating revenue amounts to obtain the comprehensive operating profit margin for each province, then identify the province with the lowest profit margin. What is the profit margin value (as a percentage, rounded to two decimal places) for the province with the lowest comprehensive operating profit margin?",
+  "guidelines": "The answer should be a percentage value with 2 decimal places. Output only the number, without the % symbol or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 0.59,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter policy records with industry field containing \"Food and Beverage Industry\" from policy_release_status.csv, 16 records in total. Group by province field to count policy numbers per province; 4 national-level policies need to be added to each province. After adding national-level policy count to each province's count, filter provinces with policy count >= 3, totaling 9 provinces: ['Yunnan', 'Sichuan', 'Ningxia', 'Hebei', 'Henan', 'Hainan', 'Hunan', 'Gansu', 'Guizhou'].",
+    "Filter all enterprise records with industry=\"Food and Beverage Industry\" from company_profile.csv, extract company name, bmCode, and province fields; 247 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract operating profit amount and operating revenue amount fields.",
+    "Filter valid enterprises with non-null operating revenue amount; 247 enterprises in total.",
+    "Group by province field, aggregate the sum of operating profit amounts and sum of operating revenue amounts for each province.",
+    "Calculate comprehensive operating profit margin for each province = sum of operating profit amounts / sum of operating revenue amounts × 100%.",
+    "Among the 9 provinces with >=3 policies, sort by operating profit margin in ascending order; the province with the lowest operating profit margin is Hainan, with total operating profit of 78,834,917.61 yuan, total operating revenue of 13,274,274,000.99 yuan, and operating profit margin = 0.59%."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Total food and beverage industry policies": 16,
+    "Provinces with >=3 policies": [
+      "Yunnan",
+      "Sichuan",
+      "Ningxia",
+      "Hebei",
+      "Henan",
+      "Hainan",
+      "Hunan",
+      "Gansu",
+      "Guizhou"
+    ],
+    "Number of valid enterprises": 247,
+    "Hainan total operating profit": 78834917.61,
+    "Hainan total operating revenue": 13274274000.99,
+    "Hainan operating profit margin (%)": 0.59
+  }
+}

assets/qa_gold/comprehensive_decision/hard008.json

@@ -0,0 +1,29 @@
+{
+  "id": "hard008",
+  "question": "In 2022, a private equity institution sought to identify high-quality provinces in the electricity, heat, gas and water production and supply industry that combine growth potential, market undervaluation, and innovation resilience. The screening logic has three layers: The first layer requires that the median year-over-year change in operating revenue of enterprises within the province be positive (>0%), to exclude regions where revenue is already shrinking; The second layer, based on the first layer results, further requires that the province's market valuation level be relatively low, i.e., the P/S ratio of all enterprises in the province must be lower than the median P/S ratio across all provinces nationwide (national median is calculated from the provincial P/S ratio series); The third layer adds an innovation requirement, i.e., the mean R&D investment ratio of enterprises in the province must be higher than the mean of all enterprises in the industry with R&D investment ratio records. How many provinces satisfy all three conditions simultaneously?",
+  "guidelines": "The answer should be an integer. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 4,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter all provincial records with industry=\"Electricity, Heat, Gas and Water Production and Supply\" from regional_industry_status.csv, extract province, median year-over-year change in operating revenue, total company market cap, total operating revenue amount, and number of enterprises; 34 records in total.",
+    "Filter enterprises with industry=\"Electricity, Heat, Gas and Water Production and Supply\" from company_profile.csv, extract company name, bmCode, and province fields; 189 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract R&D investment ratio field. 122 enterprises have non-null R&D investment ratio; national average R&D investment ratio is 1.1470%. Group by province to calculate average R&D investment ratio per province.",
+    "Filter 16 valid provinces with non-null and >0 total operating revenue amount; calculate P/S ratio per province = total company market cap / total operating revenue amount (converted to 100 million yuan); national median P/S ratio is 1.024358.",
+    "Filter high-growth provinces with median year-over-year change in operating revenue > 0%; 15 provinces in total.",
+    "Among high-growth provinces, filter low-valuation provinces with P/S ratio < national median 1.024358; 7 provinces in total.",
+    "Among high-growth, low-valuation provinces, further filter provinces where average enterprise R&D investment ratio > 1.1470%; 4 provinces ultimately satisfy all three conditions: ['Guangdong', 'Shanghai', 'Henan', 'Hebei']."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Number of valid provinces": 16,
+    "National median P/S ratio": 1.024358,
+    "National average enterprise R&D investment ratio (%)": 1.147,
+    "Number of high-growth provinces": 15,
+    "Number of high-growth, low-valuation provinces": 7,
+    "Number of provinces meeting all conditions": 4
+  }
+}

assets/qa_gold/comprehensive_decision/hard009.json

@@ -0,0 +1,31 @@
+{
+  "id": "hard009",
+  "question": "In 2022, an investment manager at a merger and acquisition fund was seeking \"high R&D, low valuation\" M&A targets in the textile, footwear and apparel industry, but the scope was limited to provinces covered by textile, footwear and apparel industry-related policies. The prerequisite for screening valid enterprises is: net profit amount strictly greater than zero, and both R&D investment ratio and company market cap fields have data records. On this basis, first use all valid enterprises in the industry as the benchmark population to calculate the median R&D investment ratio and the median P/E ratio respectively; then from the subset of valid enterprises located in policy-covered provinces, filter enterprises whose R&D investment ratio is higher than the industry median and whose P/E ratio is lower than the industry median. How many enterprises satisfy the above dual screening conditions? (P/E ratio = company market cap (100 million yuan) ÷ net profit amount (100 million yuan))",
+  "guidelines": "The answer should be an integer. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 9,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter policy records with industry field containing \"Textile, Footwear and Apparel\" from policy_release_status.csv, 21 records in total. Extract province field, remove nulls and exclude \"National\" level, to obtain 11 policy-covered provinces: ['Shanghai', 'Sichuan', 'Shandong', 'Guangdong', 'Guangxi', 'Xinjiang', 'Hebei', 'Hunan', 'Fujian', 'Liaoning', 'Shaanxi'].",
+    "Filter all enterprise records with industry=\"Textile, Footwear and Apparel\" from company_profile.csv, extract company name, bmCode, and province fields; 177 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract R&D investment ratio, net profit amount, and company market cap fields; 177 records after merge.",
+    "Filter valid enterprises with net profit amount > 0 and both R&D investment ratio and company market cap non-null; 81 enterprises in total.",
+    "Confirm data units: company market cap is in 100 million yuan, net profit amount is in yuan. Unify units for P/E calculation: P/E = company market cap (100 million yuan) ÷ (net profit amount (yuan) ÷ 100000000), i.e., P/E = company market cap (100 million yuan) ÷ net profit amount (100 million yuan).",
+    "Calculate industry-wide median benchmarks for valid enterprises: median R&D investment ratio is 2.8, median P/E = company market cap (100 million yuan) / net profit amount (100 million yuan) is 23.13.",
+    "Among valid enterprises, filter those whose province is in the policy-covered province list; 29 enterprises in total.",
+    "Among the 29 valid enterprises in policy-covered provinces, filter enterprises with R&D investment ratio > 2.8 and P/E < 23.13; 9 enterprises in total."
+  ],
+  "steps_num": 8,
+  "milestone": {
+    "Number of policy-covered provinces": 11,
+    "Total textile, footwear and apparel enterprises": 177,
+    "Number of valid enterprises": 81,
+    "Median R&D investment ratio": 2.8,
+    "Median P/E": 23.13,
+    "Valid enterprises in policy-covered provinces": 29,
+    "Number of enterprises meeting conditions": 9
+  }
+}

assets/qa_gold/comprehensive_decision/hard010.json

@@ -0,0 +1,30 @@
+{
+  "id": "hard010",
+  "question": "In 2022, to quantify the comprehensive competitive strength of the construction industry across regions, an industry association constructed a provincial competitiveness index system. The index is composed of four weighted sub-dimensions: market size share of national total (weight 30%), asset operation efficiency i.e. operating profit to total assets ratio (weight 30%), technology accumulation level i.e. cumulative invention patent grants to number of enterprises in jurisdiction ratio (weight 20%), and talent structure i.e. R&D personnel as share of total employees (weight 20%). The four raw indicators are each min-max normalized across all valid provinces, then weighted and summed to obtain the final index. Only provinces with data records for all four indicators are included in the calculation. Finally, please calculate the index difference between the first-ranked province and the last-ranked province in the competitiveness index ranking (rounded to two decimal places).",
+  "guidelines": "The answer should be a numeric value with 2 decimal places. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 0.71,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter all provincial records with industry=\"Construction\" from regional_industry_status.csv, extract province, total operating revenue amount, total operating profit amount, total assets, total cumulative Chinese invention patent grants, number of enterprises, total R&D personnel count, and total employee count fields; 34 records in total.",
+    "Filter records with industry=\"Construction\" from national_industry_status.csv to obtain national construction industry total operating revenue amount of 12,683,425,500,139.00 yuan and total number of enterprises 148.",
+    "Filter enterprises with industry=\"Construction\" from company_profile.csv, 148 enterprises in total, for cross-validation of provincial-level data.",
+    "Calculate four raw indicators per province: scale index = total operating revenue amount / 12,683,425,500,139.00, efficiency index = total operating profit amount / total assets, innovation index = total cumulative Chinese invention patent grants / number of enterprises, talent index = total R&D personnel count / total employee count. Filter 16 valid provinces with all four indicators non-null.",
+    "Apply min-max normalization to each of the four indicators: normalized value = (raw value - min) / (max - min).",
+    "Calculate competitiveness index = normalized scale index × 0.3 + normalized efficiency index × 0.3 + normalized innovation index × 0.2 + normalized talent index × 0.2.",
+    "Sort by competitiveness index in descending order; the difference between first-ranked Beijing (0.7287) and last-ranked Liaoning (0.0142) = 0.7145."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Number of valid provinces": 16,
+    "National total construction enterprises": 148,
+    "First-ranked province": "Beijing",
+    "First-ranked score": 0.7287,
+    "Last-ranked province": "Liaoning",
+    "Last-ranked score": 0.0142,
+    "Difference": 0.7145
+  }
+}

assets/qa_gold/comprehensive_decision/hard011.json

@@ -0,0 +1,50 @@
+{
+  "id": "hard011",
+  "question": "In 2022, a think tank was commissioned to study the impact of policy intervention on R&D behavior in the communication transmission equipment industry. The research design divides all enterprises with R&D investment ratio data records into two groups: one group from provinces that have appeared in policy release information with \"Communication Transmission Equipment\" related policy entries (\"National\" level entries do not count as provinces and are not included in either group); the other group from provinces that have never appeared in the above policy entries. After grouping, calculate the arithmetic mean of R&D investment ratio for each group respectively, then compute the difference between them (policy-covered provinces mean minus non-policy-covered provinces mean). This difference reflects the association between policy coverage and R&D intensity of communication transmission equipment enterprises within the jurisdiction. What is this difference in percentage points?",
+  "guidelines": "The answer should be a numeric value with 2 decimal places. A positive number indicates policy-covered provinces are higher; a negative number indicates non-policy-covered provinces are higher. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 4.9,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter policy records with industry field containing \"Communication Transmission Equipment\" from policy_release_status.csv, 70 records in total. Extract unique province values and exclude \"National\" to obtain 17 policy-covered provinces: ['Anhui', 'Shandong', 'Guangdong', 'Sichuan', 'Hubei', 'Fujian', 'Jiangxi', 'Chongqing', 'Hunan', 'Yunnan', 'Guizhou', 'Henan', 'Shaanxi', 'Hainan', 'Beijing', 'Shanghai', 'Xinjiang'].",
+    "Filter enterprise records with industry=\"Communication Transmission Equipment\" from company_profile.csv, extract company name, bmCode, and province fields; 120 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract R&D investment ratio field; 120 records after merge.",
+    "Filter valid enterprises with non-null R&D investment ratio; 117 enterprises in total.",
+    "Based on the policy-covered province list from step 1, divide valid enterprises into two groups: 86 enterprises in policy-covered provinces, 31 enterprises in non-policy-covered provinces.",
+    "Calculate mean R&D investment ratio for each group: policy-covered provinces average = 14.35%, non-policy-covered provinces average = 9.45%.",
+    "Calculate difference = policy-covered average R&D ratio - non-policy-covered average R&D ratio = 14.35 - 9.45 = 4.90 percentage points."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Number of communication transmission equipment policies": 70,
+    "Policy-covered provinces": [
+      "Anhui",
+      "Shandong",
+      "Guangdong",
+      "Sichuan",
+      "Hubei",
+      "Fujian",
+      "Jiangxi",
+      "Chongqing",
+      "Hunan",
+      "Yunnan",
+      "Guizhou",
+      "Henan",
+      "Shaanxi",
+      "Hainan",
+      "Beijing",
+      "Shanghai",
+      "Xinjiang"
+    ],
+    "Number of communication transmission equipment enterprises": 120,
+    "Number of valid enterprises": 117,
+    "Enterprises in policy-covered provinces": 86,
+    "Enterprises in non-policy-covered provinces": 31,
+    "Policy-covered average R&D ratio (%)": 14.35,
+    "Non-policy-covered average R&D ratio (%)": 9.45,
+    "Difference (percentage points)": 4.9
+  }
+}

assets/qa_gold/comprehensive_decision/hard012.json

@@ -0,0 +1,29 @@
+{
+  "id": "hard012",
+  "question": "In 2022, an antitrust research team analyzed the provincial market structure of the metal smelting and rolling processing industry. To ensure statistical reliability, only provinces with operating revenue amount records and at least 5 enterprises in the industry within the jurisdiction were included. Among qualifying provinces, the Herfindahl-Hirschman Index (HHI) was used to measure market concentration in each province: calculate each enterprise's operating revenue as a share of total operating revenue of all valid enterprises in the province, sum the squares of these shares and multiply by 100% to obtain the province's HHI value. Higher HHI indicates more concentrated markets and greater monopoly risk. After identifying the province with the highest HHI, extract the province's total operating profit amount and total operating revenue amount from provincial industry summary data, and calculate the corresponding operating profit margin. What is the operating profit margin of the province with the highest HHI?",
+  "guidelines": "The answer should be a percentage value with 2 decimal places. Output only the number, without the % symbol or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 4.14,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter enterprise records with industry=\"Metal Smelting and Rolling Processing\" from company_profile.csv, extract company name, bmCode, and province fields; 145 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract operating revenue amount field; 145 records after merge.",
+    "Filter 111 enterprises with non-null operating revenue amount; group by province to count enterprises per province; retain 13 provinces with enterprise count >= 5: ['Shanghai', 'Yunnan', 'Beijing', 'Sichuan', 'Anhui', 'Shandong', 'Guangdong', 'Jiangsu', 'Jiangxi', 'Henan', 'Zhejiang', 'Liaoning', 'Hong Kong'].",
+    "Within each valid province, calculate each enterprise's market share = enterprise operating revenue amount / sum of operating revenue amounts of all enterprises in the province.",
+    "Calculate Herfindahl-Hirschman Index (HHI) for each province = sum of squares of enterprise market shares × 100%; sort by HHI in descending order.",
+    "Among qualifying provinces, using the same valid enterprise sample as for HHI calculation (non-null operating revenue and province enterprise count ≥ 5), aggregate total operating profit amount and total operating revenue amount per province; calculate operating profit margin per province = sum of operating profit amounts / sum of operating revenue amounts × 100%.",
+    "The province with the highest HHI is Shanghai, HHI = 88.47. Total operating profit of valid enterprises in this province is 16,186,839,594.21 yuan, total operating revenue is 391,233,407,230.44 yuan; operating profit margin = 4.14%."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "Number of metal smelting and rolling processing enterprises": 145,
+    "Enterprises with non-null operating revenue": 111,
+    "Number of valid provinces (enterprises >= 5)": 13,
+    "Province with highest HHI": "Shanghai",
+    "HHI value": 88.47,
+    "Shanghai operating profit margin (%)": 4.14
+  }
+}

assets/qa_gold/comprehensive_decision/hard013.json

@@ -0,0 +1,35 @@
+{
+  "id": "hard013",
+  "question": "In 2022, a provincial industry and information department sought to evaluate the government subsidy utilization efficiency of enterprises of different sizes in the rubber and plastic products industry. After dividing enterprises into three groups by total assets—large (top 1/3 rounded up), medium (middle 1/3 rounded up), and small (bottom 1/3)—which enterprise size group has the highest subsidy utilization efficiency? What is its efficiency value?",
+  "guidelines": "The answer should be a numeric value with 2 decimal places. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": [
+    "Large",
+    166.33
+  ],
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter all enterprise records with industry=\"Rubber and Plastic Products\" from company_profile.csv, extract company name, bmCode, and province fields; 107 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract total assets, operating revenue amount, and government reward funds and subsidies fields; 107 records after merge.",
+    "Filter records with industry=\"Rubber and Plastic Products\" from national_industry_status.csv to obtain national benchmark data: total enterprises 107, total operating revenue 313,571,405,678.69 yuan, total government subsidies 1,966,394,638.00 yuan, total assets 476,387,663,666.96 yuan.",
+    "Filter valid enterprises with all three fields (total assets, operating revenue amount, government reward funds and subsidies) non-null and government reward funds and subsidies > 0; 107 enterprises in total.",
+    "Sort by total assets in descending order; divide 107 enterprises into three groups: large enterprise group (top 36, highest 1/3 by total assets), medium enterprise group (middle 36), small enterprise group (bottom 35, lowest 1/3 by total assets).",
+    "Calculate subsidy utilization efficiency per group = sum of operating revenue amounts of enterprises in the group / sum of government reward funds and subsidies of enterprises in the group. Small group efficiency = 95.58, medium group efficiency = 161.13, large group efficiency = 166.33.",
+    "The group with the highest subsidy utilization efficiency is the large enterprise group, efficiency value = 166.33."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "National total rubber and plastic enterprises": 107,
+    "Number of valid enterprises": 107,
+    "Small group enterprise count": 35,
+    "Medium group enterprise count": 36,
+    "Large group enterprise count": 36,
+    "Small group efficiency": 95.58,
+    "Medium group efficiency": 161.13,
+    "Large group efficiency": 166.33,
+    "Highest efficiency group": "Large"
+  }
+}

assets/qa_gold/comprehensive_decision/hard014.json

@@ -0,0 +1,30 @@
+{
+  "id": "hard014",
+  "question": "In 2022, a technology innovation fund evaluated the \"R&D-patent conversion\" full-chain efficiency of the consumer electronics and electrical industry across provinces, seeking to identify the province with optimal conversion efficiency (only provinces with valid enterprise count >= 3 are included). What is the R&D-patent conversion efficiency value of that province? (R&D-patent conversion efficiency = sum of annual Chinese invention patent grants / sum of annual Chinese invention patent applications × R&D output density; R&D output density = sum of annual Chinese invention patent applications / sum of R&D investment amount (100 million yuan))",
+  "guidelines": "The answer should be a numeric value with 2 decimal places. Output only the number, without units or text. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 47.29,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter all enterprise records with industry=\"Consumer Electronics and Electrical\" from company_profile.csv, extract company name, bmCode, and province fields; 358 enterprises found.",
+    "Join with company_operation_status.csv via bmCode to extract annual Chinese invention patent applications, annual Chinese invention patent grants, and R&D investment amount fields; 358 records after merge.",
+    "Filter records with industry=\"Consumer Electronics and Electrical\" from national_industry_status.csv to obtain national benchmark data: total enterprises 358, total R&D investment amount 245,156,000,000.00 yuan, total annual invention patent applications 63,940, total annual invention patent grants 43,780.",
+    "Filter 266 valid enterprises with all three fields (annual Chinese invention patent applications, annual Chinese invention patent grants, R&D investment amount) non-null; retain 13 provinces with valid enterprise count >= 3 after grouping by province.",
+    "Aggregate by province: sum of annual Chinese invention patent applications, sum of annual Chinese invention patent grants, and sum of R&D investment amount.",
+    "Calculate R&D output density per province = sum of annual Chinese invention patent applications / sum of R&D investment amount (100 million yuan); conversion efficiency = (sum of annual Chinese invention patent grants / sum of annual Chinese invention patent applications) × R&D output density.",
+    "Sort by conversion efficiency in descending order; the province with the highest conversion efficiency is Shandong, with 24,694 patent applications, 15,143 patent grants, R&D output density 77.1105, conversion efficiency = 47.29."
+  ],
+  "steps_num": 7,
+  "milestone": {
+    "National total consumer electronics and electrical enterprises": 358,
+    "Number of valid enterprises": 266,
+    "Number of valid provinces": 13,
+    "Shandong patent applications": 24694.0,
+    "Shandong patent grants": 15143.0,
+    "Shandong R&D output density": 77.1105,
+    "Shandong conversion efficiency": 47.29
+  }
+}

assets/qa_gold/comprehensive_decision/hard015.json

@@ -0,0 +1,31 @@
+{
+  "id": "hard015",
+  "question": "In 2022, a provincial government evaluated the comprehensive financial health of real estate enterprises to decide which provinces (where the province has an effective enterprise count >= 3) should face strengthened risk supervision for real estate firms. What is the health score of the province with the lowest financial health? (Financial health = Profitability score × 0.4 + Solvency score × 0.3 + Growth capability score × 0.3; Profitability is measured by the average net profit margin of enterprises in that province, where net profit margin = net profit amount / operating revenue amount; Solvency is measured as 1 − the arithmetic mean of enterprises' asset-liability ratio in that province / 100; Growth capability is measured as the median of enterprises' year-over-year change in operating revenue in that province / 100; each indicator is min-max normalized across all valid provinces before being substituted into the formula.)",
+  "guidelines": "The answer should be a numerical value with 2 decimal places. Output only the number without units or text explanation. If relevant data cannot be found, please answer \"No relevant data found\"",
+  "answer": 0.07,
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "Filter records with industry=\"房地产业\" from company_profile.csv; extract enterprise name, bmCode, and province — 295 enterprises.",
+    "Join with company_operation_status.csv on enterprise name and bmCode for year=2022; extract net profit amount, operating revenue amount, asset-liability ratio, and year-over-year change in operating revenue — 295 rows after merge.",
+    "Keep records with operating revenue amount > 0 and non-null net profit amount and asset-liability ratio — 295 rows; do not additionally exclude rows by asset-liability ratio; all participate in within-province arithmetic means and subsequent normalization.",
+    "Count enterprises by province; define valid provinces as those with effective enterprise count >= 3; in this data there are 17 valid provinces, and growth capability (median YoY operating revenue change) can be computed for all of them.",
+    "On those 17 valid provinces only, compute three raw indicators per province: Profitability = arithmetic mean over enterprises of (net profit amount / operating revenue amount); Solvency = 1 − arithmetic mean of asset-liability ratio / 100; Growth capability = median of year-over-year change in operating revenue / 100.",
+    "Apply min-max normalization to profitability, solvency, and growth capability separately across all 17 valid provinces (i.e. \"all valid provinces\" means this set only, excluding provinces with only 1–2 enterprises).",
+    "Compute financial health = normalized profitability × 0.4 + normalized solvency × 0.3 + normalized growth capability × 0.3.",
+    "Sort the 17 valid provinces by financial health ascending; the lowest is Beijing: raw profitability ≈ −1.0774, raw solvency ≈ −21.9888, raw growth capability ≈ −0.1341, health score ≈ 0.0713, rounded to two decimals as 0.07."
+  ],
+  "steps_num": 8,
+  "milestone": {
+    "real_estate_enterprises_profile_merged_2022": 295,
+    "enterprise_records_province_aggregation_no_al_ratio_exclusion": 295,
+    "valid_provinces_count_enterprises_ge_3": 17,
+    "Beijing_raw_profitability": -1.0774,
+    "Beijing_raw_solvency": -21.9888,
+    "Beijing_raw_growth_capability": -0.1341,
+    "Beijing_financial_health": 0.0713
+  }
+}

assets/qa_gold/comprehensive_decision/hard016.json

@@ -0,0 +1,35 @@
+{
+  "id": "hard016",
+  "question": "In 2022, an institutional investor plans to build an equity portfolio among Haishan Chang Industrial Equipment Company, Zhongbai Jinmao Chain Company, and Sansan Dateng Heavy Industry Company. The total portfolio weight must equal 1, the portfolio-weighted asset-liability ratio must equal exactly 45%, and the portfolio-weighted year-on-year operating revenue change must equal exactly 0%. Based on these three companies' 2022 operating data, find their portfolio weights and compute the portfolio-weighted ROE. Note: each company's asset-liability ratio is computed as total liabilities ÷ total assets × 100%.",
+  "guidelines": "Answer format: weight of Haishan Chang Industrial Equipment Company, weight of Zhongbai Jinmao Chain Company, weight of Sansan Dateng Heavy Industry Company, weighted ROE. The first three weights to four decimal places; weighted ROE to three decimal places. Output numbers and commas only, with no explanatory text. If relevant data cannot be found, answer \"No relevant data found\".",
+  "answer": [
+    0.1318,
+    0.4954,
+    0.3728,
+    10.376
+  ],
+  "metadata": {
+    "db": "bm_rag_qa",
+    "level": "hard",
+    "category": "comprehensive_decision"
+  },
+  "steps": [
+    "From company_operation_status.csv, filter by bmCode for Haishan Chang Industrial Equipment Company (100071), Zhongbai Jinmao Chain Company (100120), and Sansan Dateng Heavy Industry Company (100260) to obtain their 2022 records; 3 valid rows found. Extract total assets, total liabilities, year-on-year change in operating revenue, and net profit amount.",
+    "Compute each firm's asset-liability ratio from total assets and total liabilities using total liabilities / total assets × 100%. Results: Haishan Chang Industrial Equipment Company 86.3880%, Zhongbai Jinmao Chain Company 57.6074%, Sansan Dateng Heavy Industry Company 13.6112%.",
+    "Derive shareholder equity from total assets minus total liabilities; then compute ROE = net profit amount / shareholder equity × 100%. ROE for the three firms is 15.2002%, 11.5040%, and 7.1713%, respectively.",
+    "Let the three firms' weights be w1, w2, and w3. Set up the simultaneous equations w1+w2+w3=1, 86.3880w1+57.6074w2+13.6112w3=45, and −19.28w1+12.21w2−9.41w3=0.",
+    "Solve the system of three linear equations to obtain w1=0.13180488, w2=0.49541694, w3=0.37277818.",
+    "Compute portfolio-weighted ROE=15.2002%×0.13180488+11.5040%×0.49541694+7.1713%×0.37277818; the result is 10.376%."
+  ],
+  "steps_num": 6,
+  "milestone": {
+    "Firm count": 3,
+    "Haishan Chang Industrial Equipment Company asset-liability ratio": 86.388,
+    "Zhongbai Jinmao Chain Company asset-liability ratio": 57.6074,
+    "Sansan Dateng Heavy Industry Company asset-liability ratio": 13.6112,
+    "Haishan Chang Industrial Equipment Company weight": 0.1318,
+    "Zhongbai Jinmao Chain Company weight": 0.4954,
+    "Sansan Dateng Heavy Industry Company weight": 0.3728,
+    "Portfolio-weighted ROE": 10.376
+  }
+}