Docs: clarify parquet/json dual folders and working load_dataset examples

README.md

@@ -68,6 +68,21 @@ snapshot_download("SagheerLab/Arithmetic-Reasoning", repo_type="dataset", local_
 
 Both `JSON` and `Parquet` are provided (`parquet/train.parquet` via `default` config + `json/train.json` via `json` config) - use either.
 
+```python
+# Parquet (default) - 12.6 MB
+from datasets import load_dataset
+ds = load_dataset("SagheerLab/Arithmetic-Reasoning")  # loads parquet/train.parquet
+
+# JSON - 40 MB - use json builder via data_files (config 'json' maps to json/*.json)
+from datasets import load_dataset
+ds_json = load_dataset("SagheerLab/Arithmetic-Reasoning", data_files={
+    "train": "json/train.json",
+    "validation": "json/val.json",
+    "test": "json/test.json"
+})
+# or directly: load_dataset("json", data_files="hf://datasets/SagheerLab/Arithmetic-Reasoning/json/train.json")
+```
+
 # Dataset Summary
 
 - **Size:** 150,000 examples (135K train / 7.5K val / 7.5K test), 100% programmatically verified