README.md

metadata

language:
  - en
multilinguality:
  - monolingual
size_categories:
  - 1M<n<10M
task_categories:
  - feature-extraction
  - sentence-similarity
pretty_name: NLI for SimCSE
tags:
  - sentence-transformers
dataset_info:
  - config_name: triplet
    features:
      - name: anchor
        dtype: string
      - name: positive
        dtype: string
      - name: negative
        dtype: string
    splits:
      - name: train
        num_bytes: 51033641
        num_examples: 274951
    download_size: 33517191
    dataset_size: 51033641
  - config_name: triplet-7
    features:
      - name: anchor
        dtype: string
      - name: positive
        dtype: string
      - name: negative_1
        dtype: string
      - name: negative_2
        dtype: string
      - name: negative_3
        dtype: string
      - name: negative_4
        dtype: string
      - name: negative_5
        dtype: string
      - name: negative_6
        dtype: string
      - name: negative_7
        dtype: string
    splits:
      - name: train
        num_bytes: 129065964
        num_examples: 273540
    download_size: 87886620
    dataset_size: 129065964
  - config_name: triplet-all
    features:
      - name: anchor
        dtype: string
      - name: positive
        dtype: string
      - name: negative
        dtype: string
    splits:
      - name: train
        num_bytes: 357145333
        num_examples: 1925996
    download_size: 94616052
    dataset_size: 357145333
configs:
  - config_name: triplet
    data_files:
      - split: train
        path: triplet/train-*
  - config_name: triplet-7
    data_files:
      - split: train
        path: triplet-7/train-*
  - config_name: triplet-all
    data_files:
      - split: train
        path: triplet-all/train-*

Dataset Card for NLI for SimCSE

This is a reformatting of the NLI for SimCSE Dataset used to train the BGE-M3 model. See the full BGE-M3 dataset in Shitao/bge-m3-data. Despite being labeled as Natural Language Inference (NLI), this dataset can be used for training/finetuning an embedding model for semantic textual similarity.

Dataset Subsets

triplet subset

• Columns: "anchor", "positive", "negative"
• Column types: str, str, str
• Examples:

  {
    'anchor': 'One of our number will carry out your instructions minutely.',
    'positive': 'A member of my team will execute your orders with immense precision.',
    'negative': 'We have no one free at the moment so you have to take action yourself.'
  }
  

• Collection strategy: Reading the jsonl file in the en_NLI_data directory in Shitao/bge-m3-data and taking only the first negative.
• Deduplified: No

triplet-7 subset

• Columns: "anchor", "positive", "negative_1", "negative_2", "negative_3", "negative_4", "negative_5", "negative_6", "negative_7"
• Column types: str, str, str, str, str, str, str
• Examples:

  {
    'anchor': 'One of our number will carry out your instructions minutely.',
    'positive': 'A member of my team will execute your orders with immense precision.',
    'negative_1': 'We have no one free at the moment so you have to take action yourself.',
    'negative_2': 'A poodle is running through the grass.',
    'negative_3': 'Investment and planning are growing industries in Jamaica.',
    'negative_4': 'A bearded man is rocking out on an acoustic guitar',
    'negative_5': 'The people are sunbathing on the beach.',
    'negative_6': 'A construction worker installs a door.',
    'negative_7': 'A crowd has gathered because of a dangerous situation.'
  }
  

• Collection strategy: Reading the jsonl file in the en_NLI_data directory in Shitao/bge-m3-data and taking all samples that have 7 negatives (which is by far the majority).
• Deduplified: No

triplet-all subset

• Columns: "anchor", "positive", "negative"
• Column types: str, str, str
• Examples:

  {
    'anchor': 'One of our number will carry out your instructions minutely.',
    'positive': 'A member of my team will execute your orders with immense precision.',
    'negative': 'We have no one free at the moment so you have to take action yourself.'
  }
  

• Collection strategy: Reading the jsonl file in the en_NLI_data directory in Shitao/bge-m3-data and taking each negative, but making a separate sample with each of the negatives.
• Deduplified: No