Upload README.md with huggingface_hub

README.md

@@ -1,181 +1,24 @@
 ---
-dataset_info:
-- config_name: boolq
-  features:
-  - name: label
-    dtype: int64
-  - name: idx
-    dtype: int64
-  - name: question_en
-    dtype: string
-  - name: question_cro
-    dtype: string
-  - name: passage_en
-    dtype: string
-  - name: passage_cro
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 12507181
-    num_examples: 9427
-  - name: validation
-    num_bytes: 4280681
-    num_examples: 3270
-  - name: test
-    num_bytes: 4274921
-    num_examples: 3245
-  download_size: 13358836
-  dataset_size: 21062783
-- config_name: cb
-  features:
-  - name: label
-    dtype: int64
-  - name: idx
-    dtype: int64
-  - name: premise_en
-    dtype: string
-  - name: premise_cro
-    dtype: string
-  - name: hypothesis_en
-    dtype: string
-  - name: hypothesis_cro
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 171394
-    num_examples: 250
-  - name: validation
-    num_bytes: 43024
-    num_examples: 56
-  - name: test
-    num_bytes: 184012
-    num_examples: 250
-  download_size: 274207
-  dataset_size: 398430
-- config_name: copa
-  features:
-  - name: premise
-    dtype: string
-  - name: choice1
-    dtype: string
-  - name: choice2
-    dtype: string
-  - name: label
-    dtype:
-      class_label:
-        names:
-          '0': '0'
-          '1': '1'
-  - name: question
-    dtype: string
-  - name: idx
-    dtype: int64
-  - name: changed
-    dtype: bool
-  splits:
-  - name: train
-    num_bytes: 51741
-    num_examples: 400
-  - name: validation
-    num_bytes: 12905
-    num_examples: 100
-  - name: test
-    num_bytes: 62670
-    num_examples: 500
-  download_size: 91854
-  dataset_size: 127316
-- config_name: multirc
-  features:
-  - name: label
-    dtype: int64
-  - name: idx
-    struct:
-    - name: answer
-      dtype: int64
-    - name: paragraph
-      dtype: int64
-    - name: question
-      dtype: int64
-  - name: paragraph_en
-    dtype: string
-  - name: paragraph_cro
-    dtype: string
-  - name: question_en
-    dtype: string
-  - name: question_cro
-    dtype: string
-  - name: answer_en
-    dtype: string
-  - name: answer_cro
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 93800822
-    num_examples: 27243
-  - name: validation
-    num_bytes: 15751988
-    num_examples: 4848
-  - name: test
-    num_bytes: 30420957
-    num_examples: 9693
-  download_size: 15505344
-  dataset_size: 139973767
-- config_name: record
-  features:
-  - name: passage
-    dtype: string
-  - name: query
-    dtype: string
-  - name: entities
-    list: string
-  - name: entity_spans
-    struct:
-    - name: end
-      list: int64
-    - name: start
-      list: int64
-  - name: answers
-    list: string
-  - name: idx
-    dtype: int64
-  splits:
-  - name: train
-    num_bytes: 5779378
-    num_examples: 5607
-  - name: validation
-    num_bytes: 1932387
-    num_examples: 1869
-  - name: test
-    num_bytes: 1913715
-    num_examples: 1869
-  download_size: 6011684
-  dataset_size: 9625480
-- config_name: rte
-  features:
-  - name: label
-    dtype: int64
-  - name: idx
-    dtype: int64
-  - name: premise_en
-    dtype: string
-  - name: premise_cro
-    dtype: string
-  - name: hypothesis_en
-    dtype: string
-  - name: hypothesis_cro
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 1679810
-    num_examples: 2490
-  - name: validation
-    num_bytes: 179762
-    num_examples: 277
-  - name: test
-    num_bytes: 1926812
-    num_examples: 3000
-  download_size: 2511477
-  dataset_size: 3786384
+language:
+- hr
+language_bcp47:
+- hr-Latn
+license: mit
+task_categories:
+- text-classification
+- question-answering
+- multiple-choice
+tags:
+- superglue
+- croatian
+- south-slavic
+- nlp
+- benchmark
+- nli
+- qa
+pretty_name: Croatian SuperGLUE Benchmark
+size_categories:
+- 10K<n<100K
 configs:
 - config_name: boolq
   data_files:
@@ -209,14 +52,6 @@ configs:
     path: multirc/validation-*
   - split: test
     path: multirc/test-*
-- config_name: record
-  data_files:
-  - split: train
-    path: record/train-*
-  - split: validation
-    path: record/validation-*
-  - split: test
-    path: record/test-*
 - config_name: rte
   data_files:
   - split: train
@@ -225,4 +60,120 @@ configs:
     path: rte/validation-*
   - split: test
     path: rte/test-*
+- config_name: record
+  data_files:
+  - split: train
+    path: record/train-*
+  - split: validation
+    path: record/validation-*
+  - split: test
+    path: record/test-*
 ---
+
+# Croatian SuperGLUE Benchmark
+
+A Croatian language benchmark for evaluating language models on natural
+language understanding tasks. This dataset contains Croatian translations of
+the SuperGLUE benchmark, paired with the original English source for each row.
+COPA is the native Croatian CLASSLA dataset (not translated).
+
+Sibling datasets:
+- [`permitt/superglue-sr`](https://huggingface.co/datasets/permitt/superglue-sr) (Serbian)
+- [`permitt/superglue-mne`](https://huggingface.co/datasets/permitt/superglue-mne) (Montenegrin)
+
+**WSC**, **AX-b**, and **AX-g** are not included.
+
+## Overview
+
+| Task | Type | # Labels | Train | Validation | Test |
+|------|------|:--------:|:-----:|:----------:|:----:|
+| **BoolQ**   | QA        | 2 |  9,427 | 3,270 | 3,245 |
+| **CB**      | NLI       | 3 |    250 |    56 |   250 |
+| **COPA**    | Reasoning | 2 |    400 |   100 |   500 |
+| **MultiRC** | QA        | 2 | 27,243 | 4,848 | 9,693 |
+| **RTE**     | NLI       | 2 |  2,490 |   277 | 3,000 |
+| **ReCoRD**  | QA (span) | — |  5,607 | 1,869 | 1,869 |
+
+Test labels for the four translated classification tasks (BoolQ, CB, MultiRC,
+RTE) are copied from `permitt/superglue-sr` via a join on the SuperGLUE `idx`,
+so test evaluation works out-of-the-box. COPA ships with its own gold test
+labels from CLASSLA.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+rte = load_dataset("permitt/superglue-hr", "rte")
+
+for ex in rte["train"]:
+    premise    = ex["premise_cro"]       # Croatian premise
+    hypothesis = ex["hypothesis_cro"]    # Croatian hypothesis
+    premise_en = ex["premise_en"]        # English source
+    label      = ex["label"]             # 0=entailment, 1=not_entailment
+```
+
+## Data Format
+
+Translated tasks carry both the Croatian (`_cro`) and original English (`_en`)
+text. COPA is native Croatian and has plain field names.
+
+| Task        | Fields |
+|-------------|--------|
+| **boolq**   | `passage_cro`, `question_cro`, `passage_en`, `question_en`, `label`, `idx` |
+| **cb**      | `premise_cro`, `hypothesis_cro`, `premise_en`, `hypothesis_en`, `label`, `idx` |
+| **copa**    | `premise`, `choice1`, `choice2`, `question`, `label`, `idx`, `changed` (native Croatian) |
+| **multirc** | `paragraph_cro`, `question_cro`, `answer_cro`, `paragraph_en`, `question_en`, `answer_en`, `label`, `idx` |
+| **rte**     | `premise_cro`, `hypothesis_cro`, `premise_en`, `hypothesis_en`, `label`, `idx` |
+| **record**  | `passage`, `query`, `entities`, `entity_spans`, `answers`, `idx` (Croatian, span-selection, training-ready) |
+
+### Labels
+
+- **BoolQ, MultiRC:** `0 = False`, `1 = True`
+- **CB:** `0 = entailment`, `1 = contradiction`, `2 = neutral`
+- **RTE:** `0 = entailment`, `1 = not_entailment`
+- **COPA:** `0 = choice1`, `1 = choice2`
+
+## Notes and Caveats
+
+- **COPA** is the native CLASSLA Croatian dataset (`COPA-HR`), not a
+  translation. All splits ship with gold labels.
+- **ReCoRD** is in span-selection form. The `passage`/`query` are Croatian and
+  `entity_spans` point at positions in the Croatian passage, but the `entities`
+  and `answers` strings are preserved from the Serbian source (untranslated).
+  For a minority of rows the span boundary is also slightly misaligned with
+  respect to the Croatian token boundaries. When constructing candidates,
+  prefer `passage[start:end]` over reading from `entities`.
+- **WSC, AX-b, AX-g** are not part of this release.
+
+## Citation
+
+```bibtex
+@misc{superglue-hr-2026,
+  title  = {Croatian SuperGLUE Benchmark},
+  author = {Permitt},
+  year   = {2026},
+  publisher = {Hugging Face},
+  url = {https://huggingface.co/datasets/permitt/superglue-hr}
+}
+```
+
+## Original SuperGLUE Citation
+
+```bibtex
+@inproceedings{wang2019superglue,
+  title   = {SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems},
+  author  = {Wang, Alex and Pruksachatkun, Yada and Nangia, Nikita and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R},
+  booktitle = {Advances in Neural Information Processing Systems},
+  year    = {2019}
+}
+```
+
+## Acknowledgments
+
+- Original SuperGLUE benchmark by Wang et al.
+- COPA-HR dataset from CLASSLA.
+
+## License
+
+Released under the MIT License.