---
license: apache-2.0
task_categories:
- sentence-similarity
- text-retrieval
language:
- en
pretty_name: PaECTER Dataset
dataset_info:
- config_name: train_validation_set
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  - name: neg
    dtype: string
  splits:
  - name: train
    num_examples: 1275000
  - name: validation
    num_examples: 225000
- config_name: testset
  features:
  - name: query
    dtype: string
  - name: pos
    list: string
  - name: neg
    list: string
  splits:
  - name: test
    num_examples: 1000
configs:
- config_name: train_validation_set
  data_files:
  - split: train
    path: train_validation_set/training.jsonl
  - split: validation
    path: train_validation_set/validation.jsonl
  default: true
- config_name: testset
  data_files:
  - split: test
    path: testset/test.jsonl
---
# PaECTER Dataset

The dataset contains publication numbers of patents used to train, validate, and test our models [PaECTER](https://huggingface.co/mpi-inno-comp/paecter) and [PAT SPECTER](https://huggingface.co/mpi-inno-comp/pat_specter). These publication numbers were taken from the EPO's PATSTAT database (2023 Spring version). We used the titles and abstracts of these patents as provided in PATSTAT for training and other purposes. 

The combined training and validation dataset comprises 300,000 EPO/PCT patents as focal (query) patents. Each focal patent is associated with 5 triplets, each including one positive (pos) and one negative (neg) citation:

- Training set: Consists of 255,000 focal patents, resulting in 1,275,000 rows (5 triplets per focal patent).
- Validation set: Comprises 45,000 focal patents, resulting in 225,000 rows. 

The test dataset contains 1000 rows. Each row represents a focal patent, its 5 positive citations, and 25 randomly selected unrelated patents as negative citations.

For more details, please refer to our paper, [PaECTER: Patent-level Representation Learning using Citation-informed Transformers](https://arxiv.org/abs/2402.19411)

## Citing & Authors

```
@misc{ghosh2024paecter,
      title={PaECTER: Patent-level Representation Learning using Citation-informed Transformers}, 
      author={Mainak Ghosh and Sebastian Erhardt and Michael E. Rose and Erik Buunk and Dietmar Harhoff},
      year={2024},
      eprint={2402.19411},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}
```