Update README.md

README.md

@@ -27,3 +27,40 @@ configs:
   - split: validation
     path: data/validation-*
 ---
+# Dataset Card for AppsRetrieval-TR
+
+## Table of Contents
+- [Dataset Description](#dataset-description)
+  - [Dataset Structure](#dataset-structure)
+  - [Data Fields](#data-fields)
+- [Source Data](#source-data)
+
+## Dataset Description
+
+Thesis-Abstract-Classification-11K dataset is obtained by processing a subset of [Turkish Academic Theses](umutertugrul/turkish-academic-theses-dataset) dataset.
+
+
+### Dataset Structure
+
+The original dataset was large and examples had several `subject` fields, representing the field of the thesis.
+
+In order to construct a single-class classification problem with a reasonable data size, the following steps are carried out:
+* For each example, only the first value of `subject` field was kept as the main field of the thesis to act as a label.
+* Data points for a label with less than 60 examples were dropped, which resulted in 187 unique labels.
+* Random 60 examples for each label is selected to construct a dataset of 11,220 examples.
+
+#### Split Methodology
+* If a train-val-test split is available, we use the existing divisions as provided.
+* For datasets with a train-test split only, we create a val split from the training set, sized to match the test set, and apply this across all models.
+* In cases with a train-val split, we reassign the val set as the test split, then generate a new val split from the training data following the approach above.
+* In cases with a val-test split, we split validation into train and vad sets in 80\% and 20\% proportions, respectively.
+* When only a single combined split is present, we partition the data into train, val, and test sets in 70\%, 15\%, and 15\% proportions, respectively.
+
+### Data Fields
+
+- **text**(string) : Thesis abstract
+- **label**(string) : Field of the thesis
+
+## Source Dataset
+
+[HuggingFace](umutertugrul/turkish-academic-theses-dataset)