---
configs:
- config_name: default
  data_files:
  - split: train
    path: "minor/train.json"
  - split: test
    path: "minor/test.json"
- config_name: major
  data_files:
  - split: train
    path: "major/train.json"
  - split: test
    path: "major/test.json"
- config_name: all2023
  data_files:
  - split: val
    path: "all2023/val.json"
- config_name: all2023_v2
  data_files:
  - split: train
    path: "all2023_v2/train.json"
  - split: test
    path: "all2023_v2/test.json"
---

# arXiv Classifier Data

Usage:
```
from datasets import load_dataset, DownloadMode
# download from HuggingFace
dataset = load_dataset('mlcore/arxiv-classifier', name=<CONFIG NAME>)
# load from G2
dataset = load_dataset('/share/nikola/arxiv_classifier/data/arxiv-classifier', name=<CONFIG NAME>)
```
To force the dataset to be re-generated:
```
dataset = load_dataset('/share/nikola/arxiv_classifier/data/arxiv-classifier', name=<CONFIG NAME>, download_mode=DownloadMode.FORCE_REDOWNLOAD)
```
See: https://huggingface.co/docs/datasets/v2.20.0/en/package_reference/builder_classes#datasets.DownloadMode

Standardized terminology:
- Field: Bio/cs/physics
- Subfield: Subcategories within each
    - Primary subfield (bio., cs.LG): Given primary subfield, you can infer the field
    - Secondary subfields: Includes primary subfield, but also includes any subfields that were tagged in the paper (1-5)

Old terminology to standardized terminology translation:
- Prime category = Primary subfield
- Abstract category = secondary subfield
- Major category = field

Original data: https://www.dropbox.com/scl/fo/wwu0ifghw4sco09g67frb/h?rlkey=6ddg3yab9la3zeddvmnsfktxq&e=1&dl=0

**Minor (default)**: Dataset of papers between 2010 and 2020 (with some pre-2010 papers) with balanced primary subfields.

**Major**: Dataset of papers between 2010 and 2020 (with some pre-2010 papers) with unbalanced primary subfields to better represent the true distribution of primary categories, which is dominated by a few subfields. Note that the distribution of major subfields is still truncated.

**All 2023**: All papers published in 2023.
- Train: papers with date between January and June (inclusive)
- Test: papers with date between July and December (inclusive)

![Primary subfield distribution](figures/Primary_subfield_distribution.png)

![Secondary subfield distribution](figures/Secondary_subfield_distribution.png)

To generate subfield distribution plots:
```
python plot_subfield_distributions.py --output_path <path_to_save_plots>
```

## Set up dependencies

Get subfields to ignore and subfield aliases:
```
# in your conda env
git clone https://github.com/ag2435/arxiv-classifier-next
cd arxiv-classifier-next
conda develop .
```

## Preprocessing

Transform raw data into JSON format:
```
# major/minor cats data
python preprocess_major_minor.py -d <DATASET NAME> -s <SPLIT> -op <PATH TO SAVE PREPROCESSED DATA>
# all 2023 corpus (full text)
python preprocess_all2023_v2.py
```

Additional data checks:
```
python test_paper_id.py
```

:white_check_mark: Checked that there is no data leakage between train and test splits for each dataset config

:white_check_mark: Checked validity of arXiv identifiers for each paper

<!-- [4782 Kaggle](https://www.kaggle.com/competitions/cs-4782-2024/overview) code: [4782_preprocess_data.ipynb](4782_preprocess_data.ipynb)

[Internal Kaggle](https://www.kaggle.com/competitions/ar-xive-project-baselines/overview) code: [clean_data_for_hugging_face.ipynb](https://huggingface.co/datasets/mlcore/arxiv-classifier/blob/main/clean_data_for_hugging_face.ipynb) -->

<!-- ## Todo

- Can we incorporate additional metadata (e.g., authors)? Note: author names are contained in the full text-->