Xnhyacinth/LongWorld-Real-Workflows

LongWorld Real Workflow P6 Probe

This is a private, local-engineering release for long-context SFT experiments. It contains 542 promoted rows from 12 causally executable worlds. The release passed the p6-source-dependent-probe-12-v1 quality profile and is not a production-approved dataset.

What is in this release

• 428 train rows and 114 eval rows; split boundaries are world-level.
• 164 16K, 222 32K, and 156 exact-tokenizer-verified 64K rows.
• Four company, four research-lab, and four code-forge worlds.
• 61 unique base tasks, 24 executable proofs, 23 answer programs, and 7 verified real-source relation graphs.
• 16 real-source exact-64K rows, 8 real-source base tasks, and 3 real-source families.
• Full, counterfactual, and ordered-artifact views. Counterfactual answers are obtained by executable replay.

The real-source worlds use signed exports of public GitHub workflow records and official arXiv metadata/source archives. The source text is parsed into state, causal relations, proof requirements, and answers. Synthetic workflow events supply controlled counterfactuals and delayed effects; they are explicitly classified and are not represented as authentic source records.

Quality controls

The release gate requires strict replay, source-bundle binding, source-to-state-to-answer dependence, factual/counterfactual twins, remove-one failure, single-essential-document insufficiency, body-corruption failure, contiguous-window and sparse/dense retrieval checks, semantic-growth checks, exact 64K token verification, zero exact row duplicates, and world/domain quotas. The passing receipt is included at 05_promoted/release_gate_pass.json.

Training exports use unique ShareGPT rows. B5w.datasets.yaml applies sampler weight 2 to the single B5w.json file; no second copied weight shard is included.

Unseen evaluation status

The local-engineering split artifacts report topology/operator and source-document-family axes as ready. World/entity is only world_atomic_only, so production publication of that axis is blocked. Domain-composition is blocked because this 12-world probe has fewer than two composition groups. These limitations must be resolved before a production unseen claim.

Provenance and trust scope

This package is signed with local probe role keys and is marked local_engineering. It does not have an independently controlled KMS approval or a production trust-root pin. The OpenReview exporter and parser are implemented and tested, but live OpenReview acquisition was blocked by an upstream HTTP 403 during this run; no OpenReview response is present in these rows. arXiv and GitHub provenance is included only when the exact signed replay bundles verify.

Intended use and limitations

Use this release for controlled SFT/CPT pipeline probes, ablations, and evaluator development. Do not describe it as a production-approved corpus, an OpenReview corpus, or a general long-context benchmark. The dataset is small, English-heavy, and intentionally contains only three domains. It does not yet establish unseen domain-composition generalization or external benchmark gains.

Source documents retain their original provenance and applicable licenses. The aggregate package uses license: other because source-specific terms differ; users must review source metadata and applicable licenses before redistribution or model release. The package should remain private until that review and an independent production approval are complete.