Publish native Parquet dataset shard

README.md

@@ -28,7 +28,7 @@ configs:
   - config_name: resources
     data_files:
       - split: train
-        path: data/resources.jsonl
+        path: data/resources.parquet
 ---
 
 <p align="center">
@@ -122,7 +122,7 @@ The complete [Future Directions agenda](https://github.com/ChaoYue0307/awesome-l
 
 ## Dataset Structure
 
-The primary configuration is `resources`, with one `train` split backed by `data/resources.jsonl`. `data/resources.csv` contains the same rows for spreadsheet and dataframe workflows.
+The primary configuration is `resources`, with one `train` split backed by the native `data/resources.parquet` shard. `data/resources.jsonl` and `data/resources.csv` contain the same rows for streaming, spreadsheet, and dataframe workflows.
 
 | Field group | Fields | What it describes |
 | --- | --- | --- |

data/README.md

@@ -3,7 +3,8 @@
 Download deterministic tabular exports of all 545 source-audited resources.
 
 - `resources.csv` - Tabular export for spreadsheets and ad hoc analysis.
-- `resources.jsonl` - JSON Lines export; this file backs the Hugging Face Dataset Viewer (the dataset card's configs point at it).
+- `resources.jsonl` - JSON Lines export and source for the Hugging Face Parquet build.
+- `resources.parquet` - Native Parquet shard generated in the Hugging Face release snapshot; it powers the Dataset Viewer without relying on server-side conversion.
 - `../docs/assets/resources.json` - Slim generated payload used by the website's searchable Resource Atlas.
 - `first_seen.json` - Forward-only sidecar mapping resource URL to the date it was first added; the export left-joins it into the `date_added` column. Empty `date_added` means the entry predates per-entry tracking (started 2026-07-15).
 - `resource_source_audit.csv` - Retrieval-time audit of every row, including URL status, source title metadata, arXiv IDs, and GitHub repository stats where available.
@@ -69,6 +70,7 @@ Build the focused Hugging Face dataset card for a staging mirror with:
 
 ```sh
 python3 scripts/build_hf_card.py --output /tmp/awesome-loop-engineering-hf/README.md
+python3 scripts/build_hf_parquet.py --output /tmp/awesome-loop-engineering-hf/data/resources.parquet
 ```
 
-The card is generated from `meta/hf_card_header.yaml`, `meta/hf_card_body.md`, the current dataset, and `CITATION.cff`. The YAML front matter is intentionally Hugging Face-only and must not be added to the GitHub README.
+The card is generated from `meta/hf_card_header.yaml`, `meta/hf_card_body.md`, the current dataset, and `CITATION.cff`. The Parquet shard is a lossless derivative of `data/resources.jsonl` and requires `pyarrow`. The YAML front matter is intentionally Hugging Face-only and must not be added to the GitHub README.

data/resources.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d363995d9a67d48c8ae194ba6fb34d362b116a7391019c8c9194efe564abbbf2
+size 435293

meta/hf_card_body.md

@@ -89,7 +89,7 @@ The complete [Future Directions agenda](https://github.com/ChaoYue0307/awesome-l
 
 ## Dataset Structure
 
-The primary configuration is `resources`, with one `train` split backed by `data/resources.jsonl`. `data/resources.csv` contains the same rows for spreadsheet and dataframe workflows.
+The primary configuration is `resources`, with one `train` split backed by the native `data/resources.parquet` shard. `data/resources.jsonl` and `data/resources.csv` contain the same rows for streaming, spreadsheet, and dataframe workflows.
 
 | Field group | Fields | What it describes |
 | --- | --- | --- |

meta/hf_card_header.yaml

@@ -28,5 +28,5 @@ configs:
   - config_name: resources
     data_files:
       - split: train
-        path: data/resources.jsonl
+        path: data/resources.parquet
 ---

scripts/build_hf_parquet.py

@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+"""Build the native Parquet shard published by the Hugging Face mirror."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_INPUT = ROOT / "data" / "resources.jsonl"
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--input", type=Path, default=DEFAULT_INPUT, help="source JSON Lines export")
+    parser.add_argument("--output", type=Path, required=True, help="destination Parquet shard")
+    args = parser.parse_args()
+
+    try:
+        import pyarrow.json as arrow_json
+        import pyarrow.parquet as parquet
+    except ModuleNotFoundError:
+        print("pyarrow is required: python3 -m pip install pyarrow", file=sys.stderr)
+        return 1
+
+    expected_rows = sum(1 for line in args.input.read_text(encoding="utf-8").splitlines() if line.strip())
+    table = arrow_json.read_json(args.input)
+    if table.num_rows != expected_rows:
+        print(f"expected {expected_rows} rows, parsed {table.num_rows}", file=sys.stderr)
+        return 1
+
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    parquet.write_table(
+        table,
+        args.output,
+        compression="zstd",
+        use_dictionary=True,
+        write_statistics=True,
+    )
+
+    metadata = parquet.read_metadata(args.output)
+    if metadata.num_rows != expected_rows or metadata.num_columns != len(table.column_names):
+        print("written Parquet metadata does not match the source table", file=sys.stderr)
+        return 1
+
+    summary = {
+        "columns": metadata.num_columns,
+        "output": str(args.output),
+        "rows": metadata.num_rows,
+    }
+    print(json.dumps(summary, sort_keys=True))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())