DataWorld-Bench: English dataset card (points to github.com/Zhanli-Li/Data-World), 0-100 scoring, release notes

README.md

@@ -1,88 +1,179 @@
+---
+language:
+- zh
+task_categories:
+- question-answering
+- text-generation
+tags:
+- agent
+- benchmark
+- data-analysis
+- causal-inference
+- multi-step-reasoning
+- llm-judge
+- sandbox
+- tool-use
+size_categories:
+- n<1K
+---
+
 # DataWorld-Bench
 
-DataWorld-Bench 是 dataworld v7 全量任务集（9,692 题）经三层精选得到的 **100 题**终版基准：
-**audit 事实过滤 → taxonomy 均衡采样 → human preference 人工校验**。
-本目录即终版，50 道来自原候选（-1 题），50 道来自挑战者（-2 题）。
-
-## 筛选链路
-
-1. **Audit 过滤（事实正确性）**
-   v7 的 9,689 个 task 归并到 1,302 个源 world
-   （`Data-World/outputs/new_world/pXXXXXX/.build/audit/audit_result.json`，
-   全部 6+ 种 audit schema 统一解析，1302/1302 覆盖）。
-   仅保留 **audit 检查零失败**的 world：**852 个**（65%），
-   其余含事实错误的 world 连同其 task 整体剔除。
-
-2. **Taxonomy 均衡采样**
-   候选池在 16 个 domain 类目（`dag_factory/categories.py`）上严重偏斜
-   （7–165 个/类）。按类目分层：每类 3 个 world、两个最大类各 4 个，
-   类内 audit 0 警告 + 检查数多者优先 → **50 个 world**，
-   每个取 cause-1 + consequence-1 = **100 道候选**。
-
-3. **Human preference 校验（题目偏好）**
-   每道入选题与同路线次优候选组成 100 对二选一
-   （cause-1 vs cause-2、consequence-1 vs consequence-2），
-   标注者只凭 human preference 选择（事实性已由 audit 保证，不复核）。
-   - 自研标注系统（`dataworld-eval/labeling`，mis 认证、独占预分配、
-     超时回收）；**每对至少 2 个不同标注者**，共 15 人 + AI 评审参与
-   - 两票分歧的 47 对全部经终审决断（owner adjudication，部分由独立盲审支持）
-   - 终局：**原 -1 题胜 50 / 挑战者 -2 题胜 50**
-     （73 对多数一致 + 27 对 owner 决断），票仓可溯：
-     `labeling/votes.json` + `labeling/adjudicated.json`
-
-## 构成
-
-- 100 个任务 = 50 个 world × 2 题（cause + consequence 各 1 道胜者）
-- 16 个 domain：每类 6 题，cat13/cat15 各 8 题
-- 逐题 provenance（票数、定案方式）见 `sample_manifest.json`
-- 相关留档：`../../results/v7_audit_all.tsv`（全量 audit 通过率）、
-  `../../results/v7_perfect_balanced50x2.tsv`（候选名单）
-
-## 任务格式
-
-与 v7 全量集一致的 harbor 布局（源目录：
-`github.com/harbor-framework/harbor/datasets/dataworld_v7`）：
+**A 100-task benchmark for agentic, evidence-grounded data analysis.**
+Each task drops an agent into a self-contained business world, hands it a
+multi-system pile of raw data, and asks for a conclusion that a strict LLM judge
+then grades against a task-specific rubric.
+
+> **Code, task-generation pipeline and evaluation harness:**
+> **<https://github.com/Zhanli-Li/Data-World>**
+
+Task content (`instruction.md`, `rubric.json`, the judge prompt and the data
+itself) is written in **Chinese**; the benchmark is otherwise model-agnostic and
+can be driven by any tool-calling agent.
+
+## What is in this repository
+
+- **100 tasks** = **50 worlds × 2 tasks each** (one `cause` analysis + one
+  `consequence` analysis of the same underlying world).
+- Each task directory is a self-contained, runnable unit: instruction, data,
+  container image recipe, verifier and rubric.
+- Roughly **23.6 GB / ~50k files**, dominated by the per-task raw data
+  (transactions, ledgers, logs, tickets, sensor readings, documents, ...).
+
+## How the 100 tasks were selected
+
+The benchmark is the final slice of a much larger generated corpus (`dataworld
+v7`, 9,692 tasks over 1,302 source worlds), reduced in three stages:
+
+1. **Audit filtering (factual correctness).** Every v7 world was re-checked by
+   automated audit; only worlds with **zero failed checks** were kept —
+   **852 of 1,302 (65%)**. Worlds containing factual errors were dropped
+   together with all of their tasks.
+2. **Taxonomy-balanced sampling.** The surviving pool is heavily skewed across
+   16 domain categories (7–165 tasks per category). Stratified sampling took 3
+   worlds per category (4 for the two largest) — preferring worlds with zero
+   audit warnings and more checks — yielding **50 worlds**, and from each world
+   the `cause-1` and `consequence-1` tasks: **100 candidates**.
+3. **Human preference validation.** Each selected task was paired against the
+   runner-up from the same world (`cause-1` vs `cause-2`, `consequence-1` vs
+   `consequence-2`) and put to two-alternative human comparison (factual
+   correctness already guaranteed by the audit, so annotators judged only task
+   preference). Every pair was seen by **at least two different annotators**
+   (15 annotators in total, plus an AI reviewer); the **47** pairs with
+   disagreeing votes were all adjudicated by the dataset owner, partly backed by
+   an independent blind review. Final outcome: **50 wins for the original
+   candidates, 50 for the challengers** (73 pairs decided by majority, 27 by
+   owner adjudication).
+
+Per-task provenance (vote counts, how the decision was reached) is recorded in
+`sample_manifest.json`; annotator identities are anonymized (`annotator-01` …
+`annotator-15`).
+
+**Composition:** 16 domain categories; 6 tasks per category except the two
+largest, which have 8. Each world contributes exactly one `cause` and one
+`consequence` task.
+
+## Task layout
 
 ```
 dataworld-pXXXXXX-<cause|consequence>-<1|2>/
-├── instruction.md            # 题面 + 数据映射 + 答案路径（/app/answer.txt）
-├── task.toml                 # schema 1.1，超时/资源/artifacts
+├── instruction.md            # the problem statement, data mapping, answer path
+├── task.toml                 # schema 1.1: timeouts, resources, artifacts
 ├── environment/
-│   ├── Dockerfile            # 基础镜像 + 镜像源 + COPY data/
-│   └── data/                 # 该 world env/ 的自包含副本
+│   ├── Dockerfile            # image recipe + COPY data/
+│   └── data/                 # self-contained copy of the world's data files
 └── tests/
-    ├── test.sh               # verifier 入口
-    ├── grade.py / judge.py / judge_prompt.py
-    ├── rubric.json           # 评分细则（仅 judge 可见）
-    └── judge.env             # judge 凭据/端点（跑评测前确认网关可用）
+    ├── test.sh               # verifier entry point
+    ├── grade.py              # scoring implementation (stdlib only)
+    ├── judge.py              # judge gateway client (stdlib only)
+    ├── judge_prompt.py       # the judge prompt template
+    ├── rubric.json           # per-task rubric (visible to the judge only)
+    └── judge.env             # judge credentials/endpoint (placeholders here)
 ```
 
-判分：LLM judge 按 rubric 打分。
+The agent writes its final answer to `/app/answer.txt`; the verifier grades that
+file and emits `/logs/verifier/reward.txt` plus `/logs/verifier/details.json`.
+
+## Scoring
+
+Each task carries its own rubric, made of **positive criteria** (things a
+correct analysis must establish) and **negative criteria** (known wrong
+conclusions, or plausible-looking paths that lead to wrong numbers). The judge
+returns one verdict per criterion:
+
+- positive criterion — `satisfied` earns its points, `not_satisfied` earns 0;
+- negative criterion — `triggered` deducts its points;
+- a conclusion the report never mentions neither earns nor loses points.
+
+The raw score is then normalized against that task's own reachable bounds:
+
+```
+reward = 100 * (raw - lo) / (hi - lo)        # in [0, 100]
+```
 
-## 运行
+where `hi` is the total of the positive criteria and `lo` the total of the
+negative ones. So **all positive criteria satisfied = 100**, **all negative
+criteria triggered = 0**.
 
-```bash
-python3 eval_sweep.py --dataset datasets/DataWorld-Bench --run-dir results/<run-name> \
-    --agent-model <model> --jobs <n>
+Two edge cases matter when you aggregate results:
+
+- **A missing or empty `/app/answer.txt`** is treated as a blank submission:
+  reward is forced to `0` (the rubric floor) regardless of the rubric's negative
+  share.
+- **A grading-infrastructure failure** (judge unreachable, unparseable verdicts)
+  is *not* expressed through the score. `details.json` then carries a
+  `judge_error` key and the reward is written as `0`. Since `0` is also a
+  legitimate score, **detect failures via `judge_error`, never via the value** —
+  otherwise a broken judge silently reads as a task the model failed.
+
+Earlier reported results used a `[-1, 1]` normalization; convert with
+`r_[-1,1] = reward / 50 - 1`.
+
+## Running the benchmark
+
+Install `docker`, then for each task build the image from
+`environment/Dockerfile` (it `COPY data/ /app/data/`), run the agent inside it,
+and finally execute `tests/test.sh` — it sources `tests/judge.env`, grades
+`/app/answer.txt` and writes the reward. The full evaluation harness lives in
+[Data-World](https://github.com/Zhanli-Li/Data-World).
+
+`tests/judge.env` selects the judge gateway and model through ordinary
+environment variables:
+
+```
+ANTHROPIC_BASE_URL   # gateway base URL
+ANTHROPIC_AUTH_TOKEN # gateway token
+ANTHROPIC_MODEL      # judge model id
 ```
 
-（与 v7 全量评测同一套 `eval_sweep.py` / `agent_harness.py` 链路。）
-
-## 发布说明（HF 版）
-
-- **判分尺度：`tests/grade.py` 把每题按自身 rubric 的上下限归一后，输出 `[0, 100]` 的
-  reward**（正分项全满足 = 100，负分项全触发 = 0）。若要与论文表格里的平均分
-  （当时口径为 `[-1, 1]`）对照，换算 `r_[-1,1] = reward / 50 - 1`。
-  judge 基础设施故障不用分值表达：`details.json` 会带 `judge_error`，此时 reward 固定写 0
-  （0 同时也是合法下界，只能靠该字段区分）。
-- **代码注释与运行日志文案为英文**；`tests/judge_prompt.py` 的 `JUDGE_PROMPT`、
-  `tests/rubric.json`、`instruction.md` 等**任务内容仍为中文**（题面与判分标准本身是中文，
-  改动会改变判分行为）。
-- `environment/Dockerfile` 与 `tests/judge.env` 里的**内部镜像源与 judge 网关凭据已替换为
-  占位符**（`<base-image>` / `<your-judge-token>` / `<judge-gateway>` /
-  `<judge-model-path-or-id>`），跑评测前请填入你自己的值。
-- `sample_manifest.json` 中的内部路径与标注者用户名已脱敏（`annotator-01..15`）。
-- **布局调整（数据内容未改）**：`dataworld-p000384-cause-1` /
-  `dataworld-p000384-consequence-1` 的 `environment/data/aml/case_narratives/` 各含
-  10,530 个文件，超过仓库单目录 10,000 文件上限，故按 `shard-00 … shard-10`（每片 ≤960）
-  分片存放，**文件名与内容与平铺版逐字节一致**。
+## Release notes for this Hugging Face copy
+
+- **Placeholders instead of credentials.** `tests/judge.env` ships with
+  `<your-judge-token>` / `<judge-gateway>` / `<judge-model-path-or-id>`, and every
+  `environment/Dockerfile` builds from `<base-image>:latest` with public package
+  mirrors. Fill in your own values before running; nothing in this repository
+  grants access to any endpoint.
+- **Scoring is on the `[0, 100]` scale** described above.
+- **Code comments and log messages are in English.** The judge prompt
+  (`judge_prompt.py`), the rubrics and the instructions remain in Chinese, since
+  they are the benchmark content itself and translating them would change the
+  grading behaviour.
+- **One layout adjustment, no content change.** The two `dataworld-p000384-*`
+  tasks carry 10,530 tiny case-narrative files each, above the 10,000-files-per-
+  directory limit of the Hub. Those directories are therefore stored as
+  `case_narratives/shard-00 … shard-10` (≤960 files per shard);
+  **file names and contents are byte-identical** to the flat layout.
+- **Anonymized annotators.** Individual annotator usernames have been replaced by
+  `annotator-01` … `annotator-15`, and internal filesystem paths and gateway
+  addresses have been removed.
+- Unrelated internal build artifacts (Python `__pycache__`) are not included.
+
+## License
+
+No license file accompanies the original task corpus; add one here before
+relying on this dataset commercially.
+
+## Citation
+
+If you use DataWorld-Bench, please link
+<https://github.com/Zhanli-Li/Data-World> and this dataset card.