roberta-base + Task Transfer (NER) --> Domain-Specific QA
Objective:
This is Roberta Base without any Domain Adaptive Pretraining --> Then trained for the NER task using MIT Movie Dataset --> Then a changed head to do the SQuAD Task. This makes a QA model capable of answering questions in the movie domain, with additional information coming from a different task (NER - Task Transfer).
https://huggingface.co/thatdramebaazguy/roberta-base-MITmovie was used as the Roberta Base + NER model.
model_name = "thatdramebaazguy/roberta-base-MITmovie-squad"
pipeline(model=model_name, tokenizer=model_name, revision="v1.0", task="question-answering")
Overview
Language model: roberta-base
Language: English
Downstream-task: NER --> QA
Training data: MIT Movie, SQuADv1
Eval data: MoviesQA (From https://github.com/ibm-aur-nlp/domain-specific-QA)
Infrastructure: 4x Tesla v100
Code: See example
Hyperparameters
Num examples = 88567
Num Epochs = 3
Instantaneous batch size per device = 32
Total train batch size (w. parallel, distributed & accumulation) = 128
Performance
Eval on MoviesQA
- eval_samples = 5032
- exact_match = 55.80286
- f1 = 70.31451
Eval on SQuADv1
- exact_match = 85.6859
- f1 = 91.96064
Github Repo:
