multi_qa / README.md
dmrau's picture
Upload dataset
b0b01e2 verified
metadata
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: id
      dtype: string
    - name: content
      dtype: string
    - name: label
      sequence: string
  splits:
    - name: train
      num_bytes: 84339725.58998138
      num_examples: 453023
  download_size: 56580316
  dataset_size: 84339725.58998138

This dataset contains a combination of the following QA datasets:

  • nq open
  • msmarco 2.1 (first 100k)
  • adverserial qa
  • hotpotqa
  • wikiqa
  • sciq
  • asqa
  • triviaqa
  • wikiqa
  • freebase_qa
  • squad_v1.1

Removed all queries with tokenized length > 128 as well as all queries that had one tokenized label > 64. tokenized with gamma-2b tokenizer.

Removed all queries in msmarco with label "No Answer Present."