Add dataset

.gitattributes

@@ -53,3 +53,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.jpg filter=lfs diff=lfs merge=lfs -text
 *.jpeg filter=lfs diff=lfs merge=lfs -text
 *.webp filter=lfs diff=lfs merge=lfs -text
+arxiv.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,9 @@
 ---
 license: pddl
 ---
+
+# arXiv Figures Dataset
+
+This dataset contains image-text pairs extracted from figures from papers in the [arXiv](https://arxiv.org) repository. The dataset can be used to train [CLIP](https://arxiv.org/abs/2103.00020) models.
+
+The data consists of a sequence of papers from arXiv up until the end of 2020 stored as JSON objects in a jsonlines file. Each object contains the arXiv id of the paper (e.g. 2103.00020), the .tar file in the [arXiv S3 bucket](https://info.arxiv.org/help/bulk_data_s3.html) the paper is located in and a list of extracted figures. For each figure, the caption (encoded in unicode) and the filename of the image is given. Note that the image files themselves are not included and have to be downloaded from the S3 bucket (e.g. using the scripts provided in the [accompanying code repo](https://github.com/nopperl/clip_arxiv_pmc).

arxiv.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa7a5f5baa4b08b19424f566674bd7b62215aef3cc48420887b58528dd80d343
+size 379847746