GSoC - Demos Project: Multi-Lingual Demo for the TyDi QA Task

[RyanMullins]

Goal: Add a demo with a multilingual model on the TyDi AQ task

Branch off of dev.

• Dataset
  • Make a Dataset wrapper for the TyDi QA dataset in the ./lit_nlp/examples/datasets directory
• Model
  • Find a compatible model in Hugging Face
  • Make a Model wrapper class in the ./lit_nlp/examples/models directory
  • Add unit tests (see examples in ./lit_nlp/examples/models/*_test.py files)
• Demo
  • Make a demo file called tydi_demo.py in the ./lit_nlp/examples directory
  • Import the TyDi QA dataset
  • Import the model
  • Configure the demo layout

[aryan1107]

Some multilingual models updated:

JAX models:

• mrm8488/bert-multi-cased-finedtuned-xquad-tydiqa-goldp
• cahya/bert-base-indonesian-tydiqa

Other JAX Q/A models:

• nbroad/mt5-small-qgen

[RyanMullins]

@iftenney can you take a look at the models above? Looks like they're all PyTorch, which is complicated internally. Maybe we need to think about training our own TF2/JAX variant of mT5?