Add dataset README

README.md

@@ -1,19 +1,49 @@
 ---
-dataset_info:
-  features:
-  - name: id
-    dtype: string
-  - name: text
-    dtype: string
-  splits:
-  - name: test
-    num_bytes: 173462462
-    num_examples: 500
-  download_size: 205335379
-  dataset_size: 173462462
-configs:
-- config_name: default
-  data_files:
-  - split: test
-    path: data/test-*
+license: other
+language:
+- en
+- zh
+tags:
+- long-context
+- benchmark
+- evaluation
+- rag
+pretty_name: LongBench Pro 128k+
 ---
+
+# LongBench Pro 128k+
+
+This dataset is a filtered subset of **LongBench Pro** (`caskcsg/LongBench-Pro`).
+
+## What is included
+
+Only examples whose `token_length` field is one of:
+
+- `128k`
+- `256k`
+
+## Columns
+
+This repo keeps only:
+
+- `id`: example identifier (copied from source)
+- `text`: the original `context` field (renamed from `context` → `text`)
+
+All other fields from the source dataset are dropped.
+
+## Intended use
+
+Use this dataset when you want to benchmark long-context behavior specifically at **≥128k** length buckets, while keeping the input surface minimal (`id`, `text`).
+
+## Provenance / attribution
+
+Source dataset: `caskcsg/LongBench-Pro`.
+
+This repo contains a derived subset. Please consult the source dataset card for:
+- full task definitions
+- original annotations/fields
+- licensing/usage terms
+
+## Reproducibility
+
+The filtering logic and transformation used to build this dataset are contained in `create_dataset.py` in this repo.