README.md

metadata

language:
  - en
license: apache-2.0
size_categories:
  - 1K<n<10K
source_datasets: tau/scrolls
task_categories:
  - text2text-generation
  - summarization
tags:
  - scrolls
  - qmsum
dataset_info:
  - config_name: default
    features:
      - name: id
        dtype: string
      - name: pid
        dtype: string
      - name: input
        dtype: string
      - name: output
        dtype: string
      - name: input_token_count
        dtype: int64
      - name: output_token_count
        dtype: int64
    splits:
      - name: train
        num_bytes: 68960760
        num_examples: 1257
      - name: validation
        num_bytes: 15700972
        num_examples: 272
      - name: test
        num_bytes: 16120860
        num_examples: 281
    download_size: 42316972
    dataset_size: 100782592
  - config_name: no-prefix
    features:
      - name: id
        dtype: string
      - name: pid
        dtype: string
      - name: input
        dtype: string
      - name: output
        dtype: string
      - name: prompt
        dtype: string
    splits:
      - name: train
        num_bytes: 68944419
        num_examples: 1257
      - name: validation
        num_bytes: 15697436
        num_examples: 272
      - name: test
        num_bytes: 16117207
        num_examples: 281
    download_size: 6180898
    dataset_size: 100759062
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
  - config_name: no-prefix
    data_files:
      - split: train
        path: no-prefix/train-*
      - split: validation
        path: no-prefix/validation-*
      - split: test
        path: no-prefix/test-*

qmsum-cleaned

prefixes

It's worth noting that each "document" in input is prefixed by a question/prompt on what the model is supposed to do. You may want to explicitly handle this in some way, or prefix your models trained on this dataset.

Most frequent "prefixes" separated via sentence-splitter in the train split:

	Sentence	Count
0	Summarize the whole meeting.	121
1	Summarize the meeting	25
2	What did the team discuss about the product cost?	4
3	How did Marketing design the product evaluation?	4
4	Summarize the wrap up of the meeting.	3
5	What did the group discuss about user requirements of the new remote control?	3
6	What did the team discuss during the product evaluation?	3
7	Summarize the meeting.	2
8	Summarize what was said about digits form	2
9	What was discussed in the meeting?	2

wordcloud

Visualized as a wordcloud (train split):

token counts