Update README.md

README.md

@@ -58,3 +58,78 @@ configs:
   - split: validation
     path: data/validation-*
 ---
+
+
+# Dataset Card for arxiv_hep-th_primary Dataset
+
+## Dataset Description
+
+- **Homepage:** [Kaggle arXiv Dataset Homepage](https://www.kaggle.com/Cornell-University/arxiv)
+- **Repository:** [hepthLlama](https://github.com/Paul-Richmond/hepthLlama)
+- **Paper:** [tbd](tbd)
+- **Point of Contact:** [Paul Richmond](mailto:p.richmond@qmul.ac.uk)
+
+### Dataset Summary
+This dataset contains metadata included in arXiv submissions.
+
+## Dataset Structure
+
+An example from the dataset looks as follows:
+
+```
+{'id': '0908.2896',
+ 'submitter': 'Paul Richmond',
+ 'authors': 'Neil Lambert, Paul Richmond',
+ 'title': 'M2-Branes and Background Fields',
+ 'comments': '19 pages',
+ 'journal-ref': 'JHEP 0910:084,2009',
+ 'doi': '10.1088/1126-6708/2009/10/084',
+ 'report-no': None,
+ 'categories': 'hep-th',
+ 'license': 'http://arxiv.org/licenses/nonexclusive-distrib/1.0/',
+ 'abstract': '  We discuss the coupling of multiple M2-branes to the background 3-form and\n6-form gauge fields of eleven-dimensional supergravity, including the coupling\nof the Fermions. In particular we show in detail how a natural generalization\nof the Myers flux-terms, along with the resulting curvature of the background\nmetric, leads to mass terms in the effective field theory.\n',
+ 'versions': [{'created': 'Thu, 20 Aug 2009 14:23:37 GMT', 'version': 'v1'}],
+ 'update_date': '2009-11-09',
+ 'authors_parsed': [['Lambert', 'Neil', ''], ['Richmond', 'Paul', '']]}
+```
+
+### Languages
+
+The text in the `abstract` field of the dataset is in English, however there may be examples 
+where the abstract also contains a translation into another language. 
+
+## Dataset Creation
+
+### Curation Rationale
+The starting point was to load v193 of the Kaggle arXiv Dataset which includes arXiv submissions upto 23rd August 2024.
+The arXiv dataset contains the following data fields:
+- `id`: ArXiv ID (can be used to access the paper)
+- `submitter`: Who submitted the paper
+- `authors`: Authors of the paper
+- `title`: Title of the paper
+- `comments`: Additional info, such as number of pages and figures
+- `journal-ref`: Information about the journal the paper was published in
+- `doi`: [Digital Object Identifier](https://www.doi.org)
+- `report-no`: Report Number
+- `abstract`: The abstract of the paper
+- `categories`: Categories / tags in the ArXiv system
+
+To arrive at the arxiv_hep-th_primary dataset, the full arXiv data 
+was filtered so that only `categories` which included 'hep-th' were retained.
+This resulted in papers that were either primarily classified as 'hep-th' or appeared cross-listed.
+For this dataset, the decision was made to focus only on papers primarily classified as 'hep-th'. 
+This meant taking only those abstracts where the first characters in `categories` were 'hep-th' 
+(see [here](https://info.arxiv.org/help/arxiv_identifier_for_services.html#indications-of-classification) for more details).
+
+We also dropped entries whose `abstract` or `comments` contained the word 'Withdrawn' or 'withdrawn' and we removed the five records which appear in the repo `LLMsForHepth/arxiv_hepth_first_overfit`.
+
+In addition, we have cleaned the data appearing in `abstract` by first replacing all occurences of '\n' with a whitespace and then removing any leading and trailing whitespace.
+
+### Data splits
+
+The dataset is split into a training, validation and test set with split percentages 70%, 15% and 15%. This was done by applying `train_test_split` twice (both with `seed=42`).
+The final split sizes are as follows:
+
+| Train | Test | Validation |
+|:---:|:---:|:---:|
+|73,768 | 15,808| 15,808   |