ger-dpr-collection / README.md
samheym's picture
Update README.md
7803376 verified
metadata
dataset_info:
  features:
    - name: dataset
      dtype: string
    - name: id
      dtype: int64
    - name: query
      dtype: string
    - name: positive_ctx
      dtype: string
    - name: negative_ctx
      dtype: string
  splits:
    - name: train
      num_bytes: 35518632119.436386
      num_examples: 40520859
  download_size: 14094945761
  dataset_size: 35518632119.436386
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Description

This dataset combines and extends several prominent German-language datasets for dense passage retrieval (DPR) and cross-encoder applications. It includes:

  • deepset/germandpr
  • deutsche-telekom/wikipedia-22-12-de-dpr
  • unicamp-dl/mmarco

Contents The dataset consists of triples in the form of:

[Query, Document , Negative Context] Each triple is generated using the cross product of entries in the original datasets whenever there are multiple elements within a single field (e.g., multiple documents or negative examples associated with a query).