1

I am trying to implement a QA system using models from huggingface. One thing I do not understand is, when I don't specify which pre-trained model I am using for question-answering, is the model chosen at random?

from transformers import pipeline

# Allocate a pipeline for question-answering

question_answerer = pipeline('question-answering')
question_answerer({

     'question': 'What is the name of the repository ?',
     'context': 'Pipeline have been included in the huggingface/transformers repository'

})

Output:

{'score': 0.5135612454720828, 'start': 35, 'end': 59, 'answer': 'huggingface/transformers'}

I know how to specify a model by adding the name of the model (bert-base-uncased for example) as a model parameter, but which one is it using when you are not specifying anything? Does it use a combination of all models on huggingface? I could not find the answer.

Ashwin Geet D'Sa
  • 6,346
  • 2
  • 31
  • 59
DarknessPlusPlus
  • 543
  • 1
  • 5
  • 18

1 Answers1

1

The model is not chosen randomly. Ever task in the pipeline selects the appropriate model whichever is close to the task. A model which is closely trained on the objective of your desired task and dataset is chosen. For example, sentiment-analysis pipeline can chose the model trained on SST task.

Likewise, for question-answering, it chooses AutoModelForQuestionAnswering class with distilbert-base-cased-distilled-squad as the default model, as SQUAD dataset is associated with question answering task.

To get the list, you can look at the variable SUPPORTED_TASKS here

Ashwin Geet D'Sa
  • 6,346
  • 2
  • 31
  • 59