Can I train Question&Answering with Question-Answer pairs only?

[gabriead]

Hi Community,
is possible to use pairs of {question, answer} as training data without context? From my understanding the framework currently only works with triples as defined by the Squad dataset. What would be the appropriate input format for the training data and do I have to prepend a taskname, if I am finetuning a new task? A list of dicts like {"question":..., "answer":....} ? Many thank's in advance!

[ZanSara]

Hello @gabriead, I believe you can't train QA models without context, but I will do some more research and confirm this to you shortly.

[gabriead]

Hi @ZanSara, there is some effort going on in trying to do that (e.g. https://arxiv.org/pdf/2104.08727.pdf). Looking forward to some more feedback from your site!

[sjrl]

Hello @gabriead we do have support for running Open Domain Question Answering models which is explained in detail here: https://haystack.deepset.ai/pipeline_nodes/answer-generator Unfortunately, we do not support training these models in Haystack. You may be able to find what you are looking for in HuggingFace transformers https://github.com/huggingface/transformers/tree/main/examples/pytorch