Scaling Data Environments from Scratch

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

DataWorld-Train-2k-part1 — the training slice (part 1 of 2)

2,135 tasks in exactly the same harbor task format as Zhanli-Li/DataWorld-Bench, and graded by the same harness. One self-contained directory per task, each carrying its own copy of the world's data, nothing mounted from a sibling task.

This repository is part 1: it holds 597 of those 2,135 tasks — every task whose world id is p000499 or below. Part 2 is Zhanli-Li/DataWorld-Train-2k-part2, which holds the other 1,538 tasks (world ids p000500–p009920). The Hub caps a repository at 1,000,000 files and the slice contains 1,011,712, so it ships as two repositories split on world id: no world is split across the two, and the union is exactly the 2,135 tasks, with nothing duplicated and nothing dropped.

Format

dataworld-train-2k-part1/
├── README.md
├── sample_manifest.json      # task list + provenance (describes the full slice)
├── synthesis_data_input.json # {input, target} pairs, paths rewritten relative
├── manifest.jsonl            # per-task metadata (world / layer / kind / dag category)
├── dataworld-pXXXXXX-<cause|consequence>-<N>/   # × 589 task directories
└── _archived-tasks/          # × 8 tasks, one .tar.gz each (see below)

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) — so any one of them can be built, run and verified on its own. The three metadata files describe the full 2,135-task slice in both repositories; entries for tasks that live in the other repository resolve there. As in the benchmark, the two files that would give the answer away (WORLD.md, dag.json) are absent, and the image build asserts their absence.

Two layout adaptations (part 1 only)

All 2,135 tasks are present and no instruction, rubric or data file was edited, but two things could not be shipped verbatim here because the Hub caps a folder at 10,000 entries: 14 data folders are sharded into shard-00/, shard-01/, … subfolders (the files keep their names and bytes, only one level of path was added, and nothing in those tasks refers to the folders by name), and 8 tasks ship as _archived-tasks/<task>.tar.gz instead of directories, because there the oversized folder is referenced by path from inside the data (p000165-consequence-1..4, p000026-cause-1/2/4, p000026-consequence-3) — tar xzf one and you get the original task directory back, byte-identical.

Both adaptations concern worlds below the p000500 boundary, so part 2 ships every task as a plain directory — no shards and no _archived-tasks/ there.

Scoring

Same harness and same scale as DataWorld-Bench: reward = 100 * (raw - lo) / (hi - lo) in [0, 100], with timeout_sec = 28800.0, cpus = 8, memory_mb = 16384 and the same placeholders in tests/judge.env. Task content is not shared — the two datasets have zero tasks in common, since bench and training worlds do not intersect.

Running a task

Build environment/Dockerfile, run the agent as user agent (its report must land at /app/answer.txt), then run tests/test.sh as user verifier after filling in tests/judge.env. For the 8 tasks in _archived-tasks/, extract the tarball first so that the task directory sits next to the others.

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.