diff --git a/DATASET_CARD.md b/DATASET_CARD.md
new file mode 100644
index 0000000..98e7464
--- /dev/null
+++ b/DATASET_CARD.md
@@ -0,0 +1,172 @@
+---
+dataset_info:
+- config_name: all-samples
+  features:
+  - name: id
+    dtype: string
+  - name: title
+    dtype: string
+  - name: original_context
+    dtype: string
+  - name: question
+    dtype: string
+  - name: answers
+    struct:
+    - name: answer_start
+      sequence: int64
+    - name: text
+      sequence: string
+  - name: validation
+    dtype: string
+  - name: context
+    dtype: string
+  splits:
+  - name: train
+    num_bytes: 55875621
+    num_examples: 10001
+  download_size: 6994468
+  dataset_size: 55875621
+- config_name: default
+  features:
+  - name: id
+    dtype: string
+  - name: title
+    dtype: string
+  - name: question
+    dtype: string
+  - name: answers
+    struct:
+    - name: answer_start
+      sequence: int64
+    - name: text
+      sequence: string
+  - name: validation
+    dtype: string
+  - name: context
+    dtype: string
+  splits:
+  - name: train
+    num_bytes: 2187382
+    num_examples: 848
+  - name: val
+    num_bytes: 347924
+    num_examples: 128
+  - name: test
+    num_bytes: 2487219
+    num_examples: 1024
+  download_size: 2042697
+  dataset_size: 5022525
+- config_name: incorrect-samples
+  features:
+  - name: id
+    dtype: string
+  - name: title
+    dtype: string
+  - name: question
+    dtype: string
+  - name: answers
+    struct:
+    - name: answer_start
+      sequence: int64
+    - name: text
+      sequence: string
+  - name: validation
+    dtype: string
+  - name: context
+    dtype: string
+  splits:
+  - name: train
+    num_bytes: 6241697
+    num_examples: 2395
+  download_size: 1492991
+  dataset_size: 6241697
+configs:
+- config_name: all-samples
+  data_files:
+  - split: train
+    path: all-samples/train-*
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*
+  - split: val
+    path: data/val-*
+  - split: test
+    path: data/test-*
+- config_name: incorrect-samples
+  data_files:
+  - split: train
+    path: incorrect-samples/train-*
+license: cc-by-nc-4.0
+task_categories:
+- question-answering
+language:
+- fo
+pretty_name: FoQA
+size_categories:
+- 1K<n<10K
+---
+
+# FoQA: Faroese Question Answering Dataset
+
+## Dataset Overview
+
+**FoQA** is a Faroese extractive question answering (also known as reading
+comprehension) dataset. It consists of 2,000 question-answer-context triples, with the
+contexts coming from Faroese Wikipedia articles. The dataset has been been created in a
+two-stage process: First, 10,000 question-answer-context triples were automatically
+generated using GPT-4-turbo. These were then manually reviewed by a native Faroese
+speaker, resulting in the final 2,000 triples. For more information about the dataset
+creation, check out our [paper](missing). All data points are available, even the ones
+that were rejected, or not manually validated at all.
+
+
+## Dataset Versions
+
+We're releasing three versions of the dataset:
+
+- `default`: The default version of the dataset, with 848 training examples, 128
+  validation examples, and 1024 test examples. These are the validated examples.
+- `all-samples`: All 10,001 examples from the original dataset, including the ones that
+  were rejected or not manually validated.
+- `incorrect-samples`: The 2,395 examples that were rejected during the manual review
+  process.
+
+We acknowledge that the main use of the dataset will be the `default` version, but we
+also provide the other versions in case they are useful for research purposes.
+
+
+## Data Fields
+
+The dataset is formatted in the standard
+[SQuAD](https://huggingface.co/datasets/rajpurkar/squad_v2), with the following
+features:
+
+- `id`: The URL of the Wikipedia article the context is from.
+- `title`: The title of the Wikipedia article the context is from.
+- `question`: The question.
+- `answers`: The answer to the question, being a dictionary with the following keys:
+  - `answer_start`: A list of character offsets in the context where the answer begins.
+  - `text`: A list of the corresponding answer texts.
+- `validation`: The validation status of the data point.
+
+
+## License
+
+The dataset is licensed under the non-commercial [CC BY-NC
+4.0](https://creativecommons.org/licenses/by-nc/4.0/) license. It is non-commercial as
+it contains generated data from OpenAI models.
+
+
+## Citation
+
+If you use FoQA in your research, please cite our paper:
+
+```bibtex
+@article{simonsen2024foqa,
+  title={FoQA: A Faroese Question-Answering Dataset},
+  author={Simonsen, Annika and Nielsen, Dan Saattrup and Einarsson, Hafsteinn},
+  journal={arXiv preprint arXiv:9999.99999},
+  year={2024}
+}
+```