sjrhuschlee/deberta-v3-base-squad2-ext-v1

deberta-v3-base for Extractive QA

This is the deberta-v3-base model, fine-tuned using the SQuAD 2.0, MRQA, AdversarialQA, and SynQA datasets. It's been trained on question-answer pairs, including unanswerable questions, for the task of Extractive Question Answering.

Overview

Language model: deberta-v3-base
Language: English
Downstream-task: Extractive QA
Training data: SQuAD 2.0, MRQA, AdversarialQA, SynQA
Eval data: SQuAD 2.0
Infrastructure: 1x NVIDIA 3070

Model Usage

import torch
from transformers import(
  AutoModelForQuestionAnswering,
  AutoTokenizer,
  pipeline
)
model_name = "sjrhuschlee/deberta-v3-base-squad2-ext-v1"

# a) Using pipelines
nlp = pipeline('question-answering', model=model_name, tokenizer=model_name)
qa_input = {
'question': 'Where do I live?',
'context': 'My name is Sarah and I live in London'
}
res = nlp(qa_input)
# {'score': 0.984, 'start': 30, 'end': 37, 'answer': ' London'}

# b) Load model & tokenizer
model = AutoModelForQuestionAnswering.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

question = 'Where do I live?'
context = 'My name is Sarah and I live in London'
encoding = tokenizer(question, context, return_tensors="pt")
start_scores, end_scores = model(
  encoding["input_ids"],
  attention_mask=encoding["attention_mask"],
  return_dict=False
)

all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0].tolist())
answer_tokens = all_tokens[torch.argmax(start_scores):torch.argmax(end_scores) + 1]
answer = tokenizer.decode(tokenizer.convert_tokens_to_ids(answer_tokens))
# 'London'

Dataset Preparation

The MRQA dataset was updated to fix some errors and formatting to work with the run_qa.py example script provided in the Hugging Face Transformers library. The changes included

• Updating incorrect answer starts locations (usually off by a few characters)
• Updating the answer text to exactly match the text found in the context

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-06
• train_batch_size: 12
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 96
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3.0

Framework versions

• Transformers 4.31.0.dev0