Scaling Data Environments from Scratch

Specify the world behind the files. Generate the data. Ground the supervision.

DataWorld-Bench

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.

Format

Every task is one directory, and a directory of task directories is the dataset — nothing else has to be configured to point a harness at this repository.

DataWorld-Bench/
├── README.md
├── sample_manifest.json                          # metadata for the 100 tasks
└── dataworld-pXXXXXX-<cause|consequence>-<1|2>/  # × 100 task directories

Each task directory is an ordinary harbor task: instruction.md, task.toml, environment/ (Dockerfile + a self-contained copy of the world's data) and tests/ (verifier, judge client, rubric.json). Tasks share no state and mount nothing from a sibling task, so any one of them can be built, run and verified on its own.

• pXXXXXX is the id of the source world, cause / consequence is the task kind (every world contributes exactly one of each), and the trailing 1 / 2 is the variant index within that kind.
• Per task: 4–12 subsystems (mean ≈ 7), 1.2 MB – 6.1 GB (median ≈ 85 MB) and 8 – 10.6k files (median ≈ 31); the whole repository is ≈ 50.2k files / 50.7 GB.
• The two files that would give the answer away (WORLD.md, dag.json) are absent from the corpus and asserted absent inside the image.
• task.toml is harbor schema_version = "1.1", one envelope for all 100 tasks: [agent] timeout_sec = 28800.0 (8 h), [verifier] timeout_sec = 1200.0, [environment] cpus = 8, memory_mb = 16384, storage_mb = 10240, allow_internet = true.

Scoring

The judge returns one verdict per rubric criterion, and the raw score is normalized against that task's own reachable bounds:

reward = 100 * (raw - lo) / (hi - lo)        # in [0, 100]

All positive criteria satisfied = 100, all negative criteria triggered = 0. A missing or empty /app/answer.txt is a blank submission and scores 0; a grading-infrastructure failure also writes 0 but carries a judge_error key in details.json — detect failures via judge_error, never via the value.

Running a task

Build the image from environment/Dockerfile, run the agent as user agent (its report must land at /app/answer.txt), then run tests/test.sh as user verifier. tests/judge.env selects the judge through ordinary environment variables (ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN / ANTHROPIC_MODEL, or the JUDGE_* equivalents) and ships with placeholders. The Dockerfile builds on the public quay.io/jupyter/scipy-notebook image, pinned to a dated tag, and needs no configuration.

Links

• Project page — https://zhanli-li.github.io/Data-World/
• Paper — https://arxiv.org/abs/2610.21656
• Code, data generation pipeline and evaluation harness — https://github.com/Zhanli-Li/Data-World
• Full collection — https://huggingface.co/collections/Zhanli-Li/dataworld

License

Released under the Apache License 2.0 — see LICENSE.