Upload README.md with huggingface_hub

README.md

@@ -241,4 +241,29 @@ configs:
     path: university/validation-*
   - split: test
     path: university/test-*
+license: other
+source_datasets: []
+tags:
+- decision-model
 ---
+
+# swde
+
+SWDE (Hao et al., SIGIR 2011) attribute records over the visible text of real web pages.
+
+SWDE has 124,291 product, book, job, movie, player, restaurant, car and university pages from 80
+websites (8 verticals, 10 sites each), each page one entity, with page-level values of 3 to 5
+attributes per vertical labelled by per-site handcrafted regular expressions over DOM text nodes
+(procedural labels; ``<NULL>`` when the page has no value). Each page's HTML is converted to
+visible text (``VisibleText``) and every value is located verbatim, word-bounded, after decoding
+character references and collapsing whitespace; attributes with an unlocated value are left out,
+never turned into nulls. Pages longer than ``window`` characters are cut to one window holding
+every located value. One config per vertical; one row per page with ``fields`` =
+``{attribute: [[(start, value), ...] per distinct value]}`` (JSON, empty for a labelled null).
+SWDE has no official split: each vertical's ten sites are split by hash (8 train, 1 validation,
+1 test; ``swde_split``), so evaluation websites are unseen. Pages are template-generated, so each
+site contributes its ``pages_per_site`` lowest-hash eligible pages.
+
+Original data: [original distribution](http://web.archive.org/web/20210630013015id_/https://codeplexarchive.blob.core.windows.net/archive/projects/swde/swde.zip). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/web_extraction.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/web_extraction.py#L440).
+
+[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.