Yu-and-Ai/kingdom-dark-continent-karma

KINGDOM Dark Continent × KARMA Training Treasure Atlas

A small, metadata-only map of overlooked Hugging Face datasets at specific training and research phases. Every upstream repository is pinned to one exact 40-character Hub commit and one small evidence file SHA-256.

This atlas does:

• distinguish corpus curation, data-order dynamics, mid-training, context extension, RLVR, tool use, preference/safety/unlearning, and evaluation;
• project candidate facts as proposal-only KINGDOM graph deltas inspired by KARMA;
• keep all Dark Continent checks unknown / not_checked / false with an advisory hold recommendation;
• preserve conflict, rights, safety, consent, and execution gates;
• publish hashes and compact Dataset Viewer rows without copying upstream training examples.

It does not download upstream data, accept gated terms, deserialize artifacts, execute tools or models, train anything, modify KINGDOM, award “karma,” rank a participant, authorize Artbitrage, or crown anyone.

The hidden treasure

The sharpest lead is allenai/Dolci-RL-Zero-IF-7B. A deterministic systematic sample of 1,400 rows (100 rows at offsets 0, 1000, …, 13000) found 11 instruction bundles with more than one distinct exact paragraph-count requirement. That observation motivates a KARMA conflicts_with_source consequence before RLVR fixture admission. It is not a whole-dataset prevalence estimate, and a ground_truth field is not treated as unquestionable truth.

A second seam adds five 2025–2026 mechanisms without copying their rows:

• RewardBench2's 102 ties cases encode that valid-answer spread should remain below correct-versus-incorrect separation, without turning the score into participant rank.
• UniEdit separates counterfactual facts that should change from locality neighbors that must not change.
• OpenMathReasoning records a negative curriculum result: a recovered 137,000-question proof batch regressed the compared SFT run.
• HelpSteer3 publishes complementary preference, feedback, edit, and principle metadata while declaring that released disagreement is censored toward agreement; the catalog does not invent a universal cross-config join.
• TMax exposes all-versus-success-only terminal trajectory structure and outcome flags as inert metadata; no recovery pairing is inferred and no command is executed.

Phase cabinet

Cabinet	Examples	First allowed use
Green	DataDecide eval, Signal and Noise, fictitious TOFU controls	Pinned evaluation rows or reviewed aggregates
Green + conflict gate	Dolci RL-Zero	Constraint metadata after contradiction scanning
Amber	DataDecide recipes, Dolmino, Longmino, FineWeb, Cosmopedia, RewardBench2 ties, UniEdit, OpenMath	Manifests, taxonomies, negative results, schemas, and hashes only
Separate consent	xLAM, WildGuardMix	Public card metadata only; gate not accepted
Quarantine	HelpSteer raw chats, TMax commands, missing-license Pythia data, Python pickles	Study/card metadata only; never raw import, execution, or deserialization

The numeric rank orders research leads, not people, agents, worth, trust, or dignity.

Files

• data/phase-seeds.jsonl — eight synthetic research-phase questions.
• data/treasure-index.jsonl — 20 reviewed metadata records.
• data/proposal-index.jsonl — compact Viewer projection. Gated subjects have no full proposal artifact.
• artifacts/proposals.jsonl — closed kingdom.kg-proposal/0.1 artifacts for public subjects; download-only, not a Dataset Viewer config.
• schema/treasure-v0.1.schema.json — closed JSON Schema for treasure rows.
• provenance/source-pins.json — exact KINGDOM, Dark Continent, KARMA, and Hub source pins.
• hash-manifest.json — sorted exact byte hashes; deliberately excludes itself.

There is no standalone events.jsonl: the current package defines events only inside a proposal, so this release does not invent a second envelope.

Rights and license boundary

Apache-2.0 covers this newly authored metadata catalog, schema, and synthetic phase prompts. It does not relicense any upstream dataset or its individual content. A card-level ODC-By, CC BY, MIT, or Apache declaration does not by itself clear copyright, privacy, personality, moral, publicity, gate, or contract rights in every source item. downstream_content_rights_cleared is therefore always false in v0.1.

Raw chats, private documents, credentials, participant profiles, web text, gated rows, mutable revisions, harvested API execution, and untrusted pickle files are excluded.

npm crossover

The matching developer previews are @agenttool/dark-continent-contract@0.1.0-dev.0 and @agenttool/dark-continent-karma@0.1.0-dev.0. The npm packages provide pure, offline contract/proposal mechanics. This Hub dataset provides reviewed metadata evidence. Neither surface applies a graph change or grants authority.

Reproduce

From the matching AgentTool source revision:

cd packages/dark-continent-karma
bun run build:hf
bun run ci

The upstream metadata was observed on 2026-07-31. Re-observation requires a new pinned revision and hash; main is never evidence.