Document dataset provenance, quality and licensing limitations

README.md

@@ -1,76 +1,71 @@
 ---
-license: apache-2.0
+license: cc-by-sa-4.0
+task_categories:
+- text-generation
+- summarization
+language:
+- tr
+tags:
+- turkish
+- wikipedia
+- knowledge-extraction
+pretty_name: Turkish Wikipedia Title–Summary Pairs
+size_categories:
+- 100K<n<1M
 ---
 
----
+# Turkish Wikipedia title–summary dataset
 
-# Turkish Wikipedia Topic-to-Summary Dataset
+This is an **experimental derivative of a Turkish Wikipedia dump**, with topic titles paired with text extracted and cleaned from articles. The original project notes describe removing markup and using the first paragraph or first few sentences as the summary.
 
-This dataset consists of title–summary pairs extracted from the Turkish Wikipedia XML dump. Each entry contains a topic title as the input and a cleaned, HTML-free summary generated from the first paragraph of the corresponding article as the output.
-It is suitable for training language models, retrieval systems, and knowledge extraction tasks.
+## Data format
 
-## Format
+The repository contains `wiki_cleaned.jsonl` with fields similar to:
 
-The dataset is provided in JSONL format. Each line represents a single record:
-```
+```json
 {
-  "input": "Title",
-  "output": "A short description or summary about the given title."
+  "input": "Konu başlığı",
+  "output": "Madde metninden çıkarılmış kısa açıklama."
 }
 ```
-## Cleaning Process
-
-The following preprocessing steps were applied:
-
-Removal of HTML tags
-
-Removal of HTML entities such as &nbsp;, &amp;, &quot;
-
-Removal of template structures ({{ ... }})
-
-Removal of Infobox fields and markup
-
-Removal of wiki links ([[Page]], [[Page|Text]])
-
-Normalization of whitespace
-
-Removal of section headers (== Heading ==)
 
-Extraction of the first paragraph or first few sentences
+The Hugging Face viewer reports approximately **816,000 rows** in one train split. No independent validation or test split is provided.
 
+```python
+from datasets import load_dataset
 
-All summaries are plain text and designed to be model-friendly.
-
-Possible Use Cases
-
-Training retrieval and ranking models
-
-Topic-to-text or title-to-abstract generation
-
-Knowledge extraction and factual reasoning tasks
-
-Pretraining or finetuning LLMs on structured encyclopedic data
-
-Building question answering systems with short factual outputs
-
+ds = load_dataset(
+    "sixfingerdev/Turkish-Knowledge-extraction_topic-summary",
+    split="train",
+)
+print(ds[0]["input"])
+print(ds[0]["output"][:200])
+```
 
-## Source
+## Processing described by the original uploader
 
-The dataset is derived from the publicly available Turkish Wikipedia dump.
-It does not contain any proprietary content and follows Wikipedia's licensing terms.
+- Extract title and opening article text.
+- Strip HTML and some MediaWiki markup, including templates, infoboxes, section headings, and wiki links.
+- Normalize whitespace and text.
 
-## Size
+**The full preprocessing code, dump revision/date, and article IDs are not included**, so exact reproduction is not currently possible. Some entries may contain imperfect markup extraction, irrelevant boilerplate, duplication, or factual errors; perform your own filtering and evaluation.
 
-Format: JSONL
+## Intended uses and limitations
 
-Each record: one topic title and its cleaned summary
+Useful as exploratory data for Turkish topic-to-text generation, retrieval experiments, and evaluation of knowledge-extraction pipelines. It is **not** a verified collection of original human-written summaries or guaranteed accurate fact pairs.
 
-Dataset size will depend on the specific Wikipedia dump version used
+- No provenance URLs or article revision identifiers are present in the card.
+- No systematic quality, privacy, or contamination audit has been published.
+- Do not directly use this split as both training and evaluation data.
 
+## License and attribution
 
-## Notes
+Because the text is taken from Wikipedia, the previous `apache-2.0` label did **not** correctly convey the rights in the source articles. The dataset is now marked **CC BY-SA 4.0**, subject to the applicable Wikimedia terms.
 
-This dataset is not an official product of Wikipedia or the Wikimedia Foundation. It is a processed derivative created for research and machine learning purposes.
+- [Turkish Wikipedia](https://tr.wikipedia.org/)
+- [Wikimedia Terms of Use](https://foundation.wikimedia.org/wiki/Terms_of_Use)
+- [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)
 
+Users must credit the Wikipedia contributors, link to the relevant license, disclose modifications, and respect share-alike requirements. **Article-level attribution records are not included**, so additional source reconstruction may be needed for compliant reuse.
 
----
+This is not an official Wikimedia Foundation release.