CoolFace
Datasetpublic

jasonyandell/texas-42-joint-world-corpus-v2

Texas 42 Joint-World Corpus v2 — the clean deck Regeneration of texas-42-joint-world-corpus on the repaired world sampler, with write-time validity assertions and per-world posterior weights. Same schema, same teacher, same seed plan (minus declaration 8 — see below). Why a v2 repo The original corpus (April 2026) was generated with a world sampler whose no-candidate branch injected domino 0-0 instead of rejecting, so 27–67% of stored worlds per decision are not… See the full description on the dataset page: https://huggingface.co/datasets/jasonyandell/texas-42-joint-world-corpus-v2.

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes1.7kdownloads
Dataset Card

Texas 42 Joint-World Corpus v2 — the clean deck

Regeneration of texas-42-joint-world-corpus on the repaired world sampler, with write-time validity assertions and per-world posterior weights. Same schema, same teacher, same seed plan (minus declaration 8 — see below).

Why a v2 repo

The original corpus (April 2026) was generated with a world sampler whose no-candidate branch injected domino 0-0 instead of rejecting, so 27–67% of stored worlds per decision are not valid 28-domino deals (mk5-main#52). The sampler was repaired 2026-07-13; this repo is the corpus regenerated under it. Differences from v1 of the pipeline:

  • —Writer-side validity assertions: every stored world is asserted to be an exact cover of the decision's unseen set with correct per-seat cardinality at write time. Malformed worlds cannot fossilize silently.
  • —Per-world posterior weights (world_weights): recorded per decision.
  • —Declaration 8 (doubles-suit) purged — it models a points-bid declaration the rules reserve for nello and is unrepresentable in the engine's base rules (mk5-main#51). 9 declarations remain; original indices are kept stable (8 is skipped).
  • —Chunk manifests with SHA-256 (MANIFEST.json) — chunked, resumable, verifiable.

The original repo stays up for provenance; consumers should migrate here.

Schema, splits, generation, loading

Identical to the original repo's README except: each decision additionally carries world_weights ([M] float, normalized), and no game carries decl_id == 8. Generated with forge.eq.generate from mk5-main on CUDA (RTX 4090, Vast.ai) instead of MPS.

These files are Python pickles — only load in environments you trust.

License

MIT — feel free to use, modify, and redistribute. Attribution appreciated.