README.md

metadata

language:
  - en
size_categories:
  - 10K<n<100K
task_categories:
  - conversational
pretty_name: Doctor & Patient
dataset_info:
  features:
    - name: prompt
      dtype: string
    - name: input_ids
      sequence: int32
    - name: length
      dtype: int64
    - name: attention_mask
      sequence: int8
  splits:
    - name: train
      num_bytes: 42127351.778204426
      num_examples: 13125
    - name: test
      num_bytes: 10534245.221795576
      num_examples: 3282
  download_size: 10917910
  dataset_size: 52661597
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
tags:
  - biology
  - medical

Dataset

This is an edited and tokenized version of the MedQuad-MedicalQnADataset dataset by keivalya. The original dataset contains 16K+ questions and answers between patient and doctor, which have been converted into a full prompt to train BioGPT by Microsoft.

Tokenizer used

microsoft/BioGPT-Large (BPE tokenizer)

Full prompt

prompt = f"""You are a helpful AI Doctor who answers medical questions. Below is a question from a patient. Your task is to answer the questions as truthfully as you can.

  ### Patient:
  {sample['Question']}

  ### Doctor:
  {sample['Answer']}"""

Notes

Since bioGPT has a max input of 1024, the full prompt was truncated to stay below this limit. The truncation strategy I used made sure that only full sentences were produced.

Please note that this dataset is for research/testing only, it should not be used in a real setting or used to give medical advice to people.