Upload README.md with huggingface_hub

README.md

@@ -1,45 +1,15 @@
----
-dataset_info:
-  features:
-  - name: task
-    dtype: string
-  - name: board
-    dtype: string
-  - name: moves
-    dtype: int32
-  - name: cluster_size
-    dtype: int64
-  - name: num_walls
-    dtype: int32
-  - name: has_walls
-    dtype: bool
-  - name: solution
-    sequence: string
-  - name: messages
-    sequence:
-    - name: role
-      dtype: string
-    - name: content
-      dtype: string
-  splits:
-  - name: train
-    num_bytes: 36448916
-    num_examples: 56002
-  - name: validation
-    num_bytes: 4553339
-    num_examples: 6997
-  - name: test
-    num_bytes: 4556909
-    num_examples: 7001
-  download_size: 7694933
-  dataset_size: 45559164
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: validation
-    path: data/validation-*
-  - split: test
-    path: data/test-*
----
+# Rush Hour 6×6 — All Entries (0/1/2 Walls)
+Derived from Michael Fogleman’s full 6×6 database. See: https://www.michaelfogleman.com/rush/
+
+## Columns
+- board: 36-char row-major; 'o' empty, 'x' wall (fixed), 'A' primary car, 'B..Z' others (len 2 or 3)
+- moves: minimal #moves (a move = slide any distance along piece axis)
+- cluster_size: states in the cluster (from source)
+- num_walls / has_walls: wall metadata
+- task: "moves" or "solve"
+- solution: optimal move list (when --task solve and found)
+- messages: chat format ready for SFT (Qwen, etc.)
+
+## Notes
+- This dataset intentionally includes 0-, 1-, and 2-wall puzzles from the official file.
+- For sequence supervision, use --task solve (heavy for 6×6; prefer --limit).