PrimeIntellect/Recursive-Task-Synthesis
Recursive Task Synthesis Tasks without completed platform artifacts or with unresolved VM validation failures are temporarily excluded. exclusions.json records the exact IDs, reasons, build IDs where available, and evidence dates/runs. Exclusions affect both metadata rows and complete TAR task packages. Runtime failures are not image-build failures or proof of incorrect gold solutions. This filter does not establish that every retained task passes gold validation. Restore a task… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Recursive-Task-Synthesis.
3995
1---2license: cc-by-4.03task_categories:4- reinforcement-learning5language:6- en7tags:8- recursive-task-synthesis9- command-line10- synthetic-tasks11configs:12- config_name: default13 data_files:14 - split: train15 path: metadata/tasks.parquet16---17 18# Recursive Task Synthesis19 20Tasks without completed platform artifacts or with unresolved VM validation21failures are temporarily excluded. `exclusions.json` records the exact IDs,22reasons, build IDs where available, and evidence dates/runs. Exclusions affect23both metadata rows and complete TAR task packages. Runtime failures are not24image-build failures or proof of incorrect gold solutions. This filter does25not establish that every retained task passes gold validation.26Restore a task by verifying its platform artifacts and the failed execution27stage, removing its exclusion, and regenerating from the pinned source.28The patch evidence below also describes excluded tasks; consult the sidecar29for current membership.30 31Source: [Zhongzhi1228/Recursive-Task-Synthesis](https://huggingface.co/datasets/Zhongzhi1228/Recursive-Task-Synthesis/tree/be44f96808d5a9b599d5cb024341ff00091adeb7),32revision `be44f96808d5a9b599d5cb024341ff00091adeb7`. **36,249 rows retained; 1235 temporarily excluded.**33 34Nine Dockerfile repairs preserve the original task problems while fixing tmux35installation. A further 47 tasks have verified repairs: 42 process checks, three36reward-output paths, and two PostgreSQL startup prerequisites. These 47 tasks37scored gold 1 and no-op 0 in live checks. Three reference scripts exit nonzero38but their verifier rewards are 1; validity follows verifier reward.39 40Repairs update task packages and corresponding metadata together. Instructions41and reference solutions remain unchanged. Dockerfile-only repairs retain42`validation_status=patched_unverified`; the verified repairs use43`patched_gold_verified`. This does not claim full-corpus correctness.44 45The Docker-daemon recovery is deferred until its platform image is rebuilt and46validated; the current release does not claim that repair. No images are built47by this preparation script.48 49`patches.json` records source hashes, patch evidence and the content-hash algorithm50used for modified tasks. Unmodified hashes retain their upstream meaning. TAR51sizes/checksums and task byte counts are updated. Preparation code lives in52[prime-data](https://github.com/PrimeIntellect-ai/prime-data/blob/fix/rts-verified-recoveries/datasets/recursive-task-synthesis/prepare.py).53Original attribution and CC-BY-4.0 license follow below.54 55---56 57This dataset contains 37,484 validated command-line task instances produced58through recursive task synthesis. Public identifiers are opaque and stable.59 60- `metadata/tasks.parquet`: one searchable row per task instance.61- `metadata/shard_manifest.jsonl`: TAR sizes and SHA256 checksums.62- `data/tasks-*.tar`: sanitized runnable task packages.63 64The searchable task rows include:65 66- `instruction`: contents of `instruction.md`.67- `task_toml`: contents of `task.toml`.68- `solution`: contents of `solution/solve.sh`, or null when absent.69- `dockerfile`: contents of `environment/Dockerfile`, or null when absent.70- `member_prefix` and `shard`: locations of the complete runnable task package.71 72```python73from datasets import load_dataset74 75tasks = load_dataset("Zhongzhi1228/Recursive-Task-Synthesis", split="train")76```77 78## License, attribution, and modifications79 80This dataset is released under the81[Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/).82See `LICENSE` for the complete legal text.83 84The dataset was produced through recursive task synthesis and executable85validation. Depending on the task, instructions, solutions, verifiers,86environments, metadata, identifiers, and packaging were rewritten, extended,87sanitized, or regenerated. Each published task is paired with searchable88metadata and a complete runnable package.89 90This license covers the original contributions and91adaptations that the publisher has authority to license. It does not grant92rights over third-party materials beyond what is permitted by their original93terms.94 95Detailed methodology, provenance, and related-work attribution are documented96in the accompanying paper.97 