Add dataset card for strict shuffled dataset

README.md

@@ -0,0 +1,25 @@
+---
+pretty_name: test strict shuffled
+task_categories:
+- text-classification
+tags:
+- parquet
+- shuffled
+- exact-shuffle
+---
+
+# Strict shuffled copy of `KantaHayashiAI/test`
+
+This dataset was produced by assigning each source row a deterministic pseudo-random sort key
+derived from:
+
+- source parquet path
+- row index inside that parquet
+- shuffle seed `2026-04-01-strict-shuffle-v1`
+
+Rows were first partitioned into `512` buckets by the high bits of the key,
+then each bucket was fully sorted by `(__key_hi, __key_lo, __file_id, __row_idx)`.
+This yields a deterministic global shuffled order without requiring the full dataset
+to be materialized twice on local disk.
+
+Expected train shard count: `512`.