| --- |
| language: en |
| license: mit |
| task_categories: |
| - text-generation |
| tags: |
| - eleusis |
| - inductive-reasoning |
| - long-horizon |
| - verifiers |
| - card-game |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # Eleusis Scaled Bank (v3, concept-balanced) |
|
|
| A 1,660-rule bank for the [Eleusis](https://huggingface.co/datasets/nph4rd/eleusis-calibrated) |
| long-horizon inductive-reasoning environment. Schema matches the default Eleusis |
| convention: `train` / `test` splits with columns `rule_id`, `label`, `family`, `code`. |
|
|
| ## Diversity model |
|
|
| Diversity here means distinct DISCOVERY PROBLEMS, not just distinct rule strings. |
| Every item belongs to a concept (a template skeleton); the test split is capped |
| at a handful of instances per concept so it measures breadth of induction, while |
| the train split keeps all instances (parameter variation prevents answer |
| memorization during training). |
|
|
| - `test` (218): ~150 distinct concepts - 36 hand-authored new concepts plus |
| <=2 instances per legacy concept. |
| - `train` (1,442): full instance pool including easy anchors and deep items. |
|
|
| ## The 14 new hand-authored concepts |
|
|
| Designed against measured GLM-5.2 difficulty ladders (482 episodes) to target |
| the 26-70 solve-turn band, exploiting what measurably makes this model struggle: |
| sparse-evidence constraints, hidden state, cross-lag relations, asymmetric gating. |
|
|
| | concept | idea | |
| |---|---| |
| | dual_window_agreement | consistent with both lag-1 and lag-2 ranks | |
| | cross_lag_sandwich | match attribute A two back, differ on B one back | |
| | streak_cap | never K consecutive accepted cards sharing an attribute | |
| | bounded_color_discrepancy | running red/black counter with a lead bound | |
| | rank_momentum_zigzag | rank deltas must alternate in sign | |
| | event_trigger | constraint fires only after face cards | |
| | rank_scarcity_window | no rank repeats within w accepted cards | |
| | alternation_escape_hatch | color alternation with a rare escape hatch | |
| | midpoint_corridor | rank near the mean of the last two ranks | |
| | positional_roles | slot index dictates which constraint applies | |
| | suit_keyed_monotone | monotone only within the same suit | |
| | cross_lag_disjunction | OR between a lag-1 and a lag-3 fact | |
| | event_flip_state | each event flips which value is acceptable | |
| | count_latch | after N reds, reds are banned | |
| |
| ## Construction & validation |
| |
| Sources: combinatorial template grids aimed at measured ladder positions, |
| single-knob mutations of 283 measured parent skeletons, unpublished earlier |
| candidates, and the hand-authored concepts above. All rows pass restricted-AST |
| compilation, starter existence, acceptance-rate bounds (trajectory simulation |
| for history-dependent rules), and behavioral uniqueness via deterministic probe |
| signatures against every published Eleusis bank. |
| |
| ## Caveats |
| |
| **No item in this dataset has been evaluated on any reference model yet.** |
| Difficulty priors are approximate (parent/template inheritance; observed GLM-5.2 |
| single-deal scatter is +/-12 turns with occasional solve/abandon flips). A final |
| frozen-protocol evaluation is planned to measure aggregate and per-item behavior |
| before benchmark use. Lineage: `construction_metadata.jsonl`. |
|
|
