mining_demo / README.md
tomaarsen's picture
tomaarsen HF staff
Librarian Bot: Add language metadata for dataset (#2)
2780cbd verified
metadata
language:
  - en
dataset_info:
  features:
    - name: query
      dtype: string
    - name: answer
      dtype: string
  splits:
    - name: train
      num_bytes: 30286
      num_examples: 100
  download_size: 24527
  dataset_size: 30286
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
tags:
  - sentence-transformers

Dataset Card for gooaq with mined hard negatives

This dataset is a collection of query-answer-negative triplets from the gooaq dataset. See gooaq for additional information. This dataset can be used directly with Sentence Transformers to train embedding models.

Mining Parameters

The negative samples have been mined using the following parameters:

  • range_min: 10, i.e. we skip the 10 most similar samples
  • range_max: 20, i.e. we only look at the top 20 most similar samples
  • margin: 0.1, i.e. we require negative similarity + margin < positive similarity, so negative samples can't be more similar than the known true answer
  • sampling_strategy: random, i.e. whether to randomly sample from the candidate negatives or take the "top" negatives
  • num_negatives: 3, i.e. we mine 3 negatives per question-answer pair

Dataset Format

  • Columns: query, answer, negative
  • Column types: str, str, str
  • Example:
{
    "query": "is toprol xl the same as metoprolol?",
    "answer": "Metoprolol succinate is also known by the brand name Toprol XL. It is the extended-release form of metoprolol. Metoprolol succinate is approved to treat high blood pressure, chronic chest pain, and congestive heart failure."
}