README.md

---
license: cc-by-4.0
task_categories:
- reinforcement-learning
language:
- en
tags:
- recursive-task-synthesis
- command-line
- synthetic-tasks
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata/tasks.parquet
---

# Recursive Task Synthesis

Only tasks with completed PUBLIC platform container and VM artifacts at the
2026-09-17T02:53:50.654736+00:00 registry audit are included. The excluded IDs, reasons
and last failed build IDs are recorded in `exclusions.json`. Exclusions affect
both metadata rows and complete TAR task packages. This is a hosting filter,
not gold validation. Recheck both artifact types and remove recovered IDs from
the preparation sidecar to restore them from the pinned source.
The patch evidence below also describes excluded tasks; consult the sidecar
for current membership.

Source: [Zhongzhi1228/Recursive-Task-Synthesis](https://huggingface.co/datasets/Zhongzhi1228/Recursive-Task-Synthesis/tree/be44f96808d5a9b599d5cb024341ff00091adeb7),
revision `be44f96808d5a9b599d5cb024341ff00091adeb7`. **37,353 rows retained; 131 temporarily excluded.**

Three Ubuntu Dockerfiles install `tmux` before the deliberately incorrect Docker
APT source. Six CentOS Dockerfiles use temporary Vault repository options for the
appended tmux installation, preserving on-disk repository configuration. Both the `dockerfile`
column and the corresponding TAR members are patched. Task IDs, instructions,
solutions, verifiers and all other task files are preserved. Changed rows use
`validation_status=patched_unverified`: upstream validation is not a new gold pass.

Two patched images (`1f90e8…`, `f6cb7d…`) built locally and passed setup checks.
Both fail untouched task assertions (3 failed, 1 passed each). The second task's
original verifier cannot bootstrap without curl; its unchanged assertions were
checked using an independently supplied runner. This verifier defect is retained
and recorded in `patches.json`. All three Ubuntu platform image builds completed.
The CentOS `256702…` local build passed and preserved original repo-file hashes;
all six CentOS platform builds are pending. No gold-solution validation or
full-corpus validation is claimed.

The 115-image FASM cluster is not excluded: its sampled Dockerfile built locally,
so remote VM-build failures remain undiagnosed. Publishing/retrying images is a
separate follow-up; this dataset does not assert complete platform image coverage.

`patches.json` records source hashes, patch evidence and the content-hash algorithm
used for modified tasks. Unmodified hashes retain their upstream meaning. TAR
sizes/checksums and task byte counts are updated. Preparation code lives in
[prime-data](https://github.com/PrimeIntellect-ai/prime-data/blob/fix/rts-dockerfile-mirror/datasets/recursive-task-synthesis/prepare.py).
Original attribution and CC-BY-4.0 license follow below.

---

This dataset contains 37,484 validated command-line task instances produced
through recursive task synthesis. Public identifiers are opaque and stable.

- `metadata/tasks.parquet`: one searchable row per task instance.
- `metadata/shard_manifest.jsonl`: TAR sizes and SHA256 checksums.
- `data/tasks-*.tar`: sanitized runnable task packages.

The searchable task rows include:

- `instruction`: contents of `instruction.md`.
- `task_toml`: contents of `task.toml`.
- `solution`: contents of `solution/solve.sh`, or null when absent.
- `dockerfile`: contents of `environment/Dockerfile`, or null when absent.
- `member_prefix` and `shard`: locations of the complete runnable task package.

```python
from datasets import load_dataset

tasks = load_dataset("Zhongzhi1228/Recursive-Task-Synthesis", split="train")
```

## License, attribution, and modifications

This dataset is released under the
[Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/).
See `LICENSE` for the complete legal text.

The dataset was produced through recursive task synthesis and executable
validation. Depending on the task, instructions, solutions, verifiers,
environments, metadata, identifiers, and packaging were rewritten, extended,
sanitized, or regenerated. Each published task is paired with searchable
metadata and a complete runnable package.

This license covers the original contributions and
adaptations that the publisher has authority to license. It does not grant
rights over third-party materials beyond what is permitted by their original
terms.

Detailed methodology, provenance, and related-work attribution are documented
in the accompanying paper.