Add dataset card (README.md)

README.md

@@ -1,23 +1,68 @@
 ---
-dataset_info:
-  features:
-  - name: wikipedia_id
-    dtype: string
-  - name: wikipedia_title
-    dtype: string
-  - name: text
-    dtype: string
-  - name: _id
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 23179414111
-    num_examples: 111789997
-  download_size: 10596182407
-  dataset_size: 23179414111
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+language:
+- en
+license: cc-by-sa-4.0
+pretty_name: KILT Wikipedia knowledge source (paragraph-level)
+tags:
+- wikipedia
+- kilt
+- retrieval
+- en
+- paragraph
+task_categories:
+- text-retrieval
 ---
+
+# KILT Wikipedia — paragraph-level (flattened)
+
+## Dataset summary
+
+This dataset is a **flattened** view of the [KILT knowledge source](https://github.com/facebookresearch/KILT) (`kilt_knowledgesource.json`): each row is **one Wikipedia paragraph** (one string in the original per-page `text` list), not one page per row.
+
+- **Source corpus:** KILT Wikipedia knowledge source (2019/08/01 Wikipedia dump, per KILT README).
+- **Split:** `train` only (full paragraph stream).
+- **Rows (train):** `111,789,997`
+- **Input shards used:** `112` JSONL file(s) (`part-*.jsonl`).
+- **Parquet chunks before Hub push:** `2236` (batch_rows=50000, Hub max_shard_size=500MB).
+
+## Data fields
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `wikipedia_id` | string | KILT Wikipedia page id |
+| `wikipedia_title` | string | Page title |
+| `text` | string | Single paragraph body |
+| `_id` | string | Stable id: `{<page _id>}::p{<paragraph_index>}` |
+
+## How it was built
+
+1. Convert `kilt_knowledgesource.json` (JSONL, one JSON object per line) with `OSCAR_like_experiments/scripts/convert_kilt_knowledge_source_to_paragraph_jsonl.py`.
+2. Upload with `OSCAR_like_experiments/scripts/push_kilt_paragraph_jsonl_to_hub.py` from directory: `/home/jovyan/rpt/OSCAR_like_experiments/scripts/kilt_knowledgesource`.
+
+## Intended use
+
+Sparse / dense **retrieval indexing** (e.g. BM25, SPLADE) where each document unit is a paragraph, matching the KILT-style chunking used in RAG pipelines.
+
+## Limitations
+
+- Text is **English Wikipedia** as packaged in KILT; formatting/markup follows KILT preprocessing.
+- Not an official Meta/Facebook dataset release; this is a **derived** redistribution — comply with Wikipedia and KILT terms.
+
+## Citation
+
+Please cite **KILT** (and Wikipedia as appropriate):
+
+```bibtex
+@inproceedings{petroni-etal-2021-kilt,
+  title = {KILT}: a Benchmark for Knowledge Intensive Language Tasks,
+  author = {Petroni, Fabio and Piktus, Aleksandra and Fan, Angela and others},
+  booktitle = {NAACL-HLT},
+  year = {2021},
+}
+```
+
+Repository: [facebookresearch/KILT](https://github.com/facebookresearch/KILT)
+
+## Dataset card contact
+
+Dataset repo: `s-nlp/kilt`