Dataset card: document zero train/test leakage

README.md

@@ -1,36 +1,55 @@
 ---
-dataset_info:
-  features:
-  - name: task
-    dtype: string
-  - name: state
-    dtype: string
-  - name: instructions
-    dtype: string
-  - name: criteria_keys
-    list: string
-  - name: criteria_values
-    list: string
-  - name: label_alias
-    dtype: string
-  - name: label_key
-    dtype: string
-  - name: meta
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 182520512
-    num_examples: 240800
-  - name: test
-    num_bytes: 3582833
-    num_examples: 4700
-  download_size: 66376464
-  dataset_size: 186103345
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: test
-    path: data/test-*
+license: apache-2.0
+pretty_name: systemone-lite Phase 2 distill
+task_categories:
+  - text-classification
+  - multiple-choice
+tags:
+  - system-one
+  - alias-ce
+  - spatial
+  - cloze
+  - zero-leakage
+size_categories:
+  - 100K<n<1M
 ---
+
+# systemone-lite-phase2
+
+Typed System One distill rows (`task` / `state` / `instructions` / criteria /
+`label_alias`) for [`systemone-lite`](https://github.com/fritzprix/systemone-lite).
+
+## Critical: train / test hygiene (2026-09-22)
+
+Earlier local mixes had **severe train∩eval state leakage** (debate ~91%,
+word_games ~87%, connect4 ~37% state_task overlap). This Hub revision is rebuilt
+with **0.00%** train∩test overlap on `state_task` fingerprints
+(`scripts/audit_train_eval_overlap.py`).
+
+| Mechanism | Detail |
+|---|---|
+| `word_games` | Disjoint TRAIN / EVAL vocabularies |
+| `debate_judge` | Held-out topics (eval never sees train motions) |
+| Spatial / chess / CA | Eval states rejected against train state hashes |
+| `nlp_cloze` | WikiText-2 **document-disjoint** (train split → train, test split → eval) |
+
+**Do not mix `train` rows into `test` (or vice versa) when building local evals.**
+
+## Splits
+
+| Split | Rows | Role |
+|---|---:|---|
+| `train` | 240800 | SFT |
+| `test` | 4700 | Held-out eval only |
+
+Gym counts (train): `{"connect4": 35000, "game2048": 35000, "gridworld": 35000, "sokoban": 35000, "chess": 32400, "cellular_automata": 12000, "nlp_cloze": 12000, "word_games": 12000, "debate_judge": 10800, "resource_allocator": 10800, "ticket_dungeon": 10800}`
+
+## Schema
+
+Flat columns for Hub friendliness: `state` and `meta` are JSON strings;
+`criteria_keys` / `criteria_values` parallel lists.
+
+## Cite / notes
+
+- Leakage + cloze write-up: repo `docs/NOTE_S1_CLOZE_AND_LEAKAGE_2026-09-22.md`
+- Not affiliated with TypeSafe AI / JevBench official leaderboards.